CN108280757A - User credit appraisal procedure and device - Google Patents

Abstract

The invention discloses a kind of user credit appraisal procedure and devices, belong to information security field.The method includes：Obtain the second behavior sequence of the first behavior sequence and at least one association user of user；By the first behavior sequence inputting first circulation neural network of user, the First ray character representation of user is obtained；Second behavior sequence of association user is inputted into second circulation neural network, the second sequence signature for obtaining association user indicates；First ray character representation and the second sequence signature are indicated into input stack self-encoding encoder, the assemblage characteristic for obtaining user derives expression；The assemblage characteristic of user is derived and indicates input grader, obtains the credit evaluation data of user.The present invention carries out feature extraction automatically and feature combination derives, it solves the problems, such as that artificial extraction feature and engineer combine derivative feature and consider that not comprehensive, efficiency is low, calculating is of high cost, has achieved the effect that feature covering is comprehensive, improves efficiency, reduced calculating cost.

Description

User credit appraisal procedure and device

Technical field

The present embodiments relate to information security field, more particularly to a kind of user credit appraisal procedure and device.

Background technology

Reference model is the data model of the credit scoring for calculating user according to the behavioral data of user.Reference model The credit evaluation being widely used in internet finance.

One reference model of structure mainly undergoes four-stage：(1) foundation characteristic extracts；(2) feature combination derives；(3) Reference model construction；(4) modelling effect is verified.Foundation characteristic is extracted for extraction and credit evaluation from the behavioral data of user Related feature；The derivative feature for foundation characteristic extraction stage to be drawn into of feature combination is combined, and it is special to obtain combination Sign is derivative to be indicated；Reference model construction, which is used to combine the assemblage characteristic that the derivative stage obtains according to feature to derive, indicates structure reference Model；Verification of the modelling effect verification for carrying out using effect to the reference model of structure.Wherein, foundation characteristic extraction and spy Sign combination is derivative to be referred to as Feature Engineering, and Feature Engineering usually requires manually to participate in completing.Foundation characteristic extraction is typically to pass through Feature related with credit evaluation is manually extracted from the behavioral data of user, such as：Extract the day of user's certain behavior generation Number, number, frequency etc..The derivative stage is combined in feature, since grader used in existing reference model belongs to line mostly Property grader, linear classifier cannot be automatically captured the interactive relation between feature, therefore be input to the feature in grader It needs by being manually combined.

Due to there is the feature that can not definitely measure in the behavioral data of user, this category feature is referred to as hidden feature, hidden Containing there may be incidence relations between feature；And hidden feature can not be extracted based on artificial foundation characteristic extraction, and also without Method utilizes the incidence relation between hidden feature, incomplete to lead to the problem of foundation characteristic extraction.In addition, due to manually setting Meter combination derivative feature is typically to rely on the priori of people, it is also possible to it is incomplete to lead to the problem of consideration.And with basis The increase of feature quantity, feature combination scale also can rapid growth, manually carry out foundation characteristic extract and feature combination spreads out The rate that comes into force is low, it is of high cost to calculate.

Invention content

In order to solve to combine the considerations of derivative feature causes not by artificial extraction feature and engineer in the prior art Comprehensively, efficiency is low, calculates problem of high cost, and an embodiment of the present invention provides a kind of user credit appraisal procedure and devices.Institute It is as follows to state technical solution：

In a first aspect, a kind of user credit appraisal procedure is provided, the method includes：

The second behavior sequence of the first behavior sequence and at least one association user of user is obtained, the association user is In social networks, there are associated other users with the user；

By the first behavior sequence inputting first circulation neural network of the user, the First ray for obtaining the user is special Sign indicates；

Second behavior sequence of the association user is inputted into second circulation neural network, obtains the of the association user Two sequence signatures indicate；

The First ray character representation and second sequence signature are indicated into input stack self-encoding encoder, obtained described The assemblage characteristic of user, which derives, to be indicated；

The assemblage characteristic of the user is derived and indicates input grader, obtains the credit evaluation data of the user；

Wherein, in the first circulation neural network, the second circulation neural network and the stack self-encoding encoder Model parameter is determined after being trained using sample sequence data, the first circulation neural network and the second circulation Model parameter in neural network is identical.

Second aspect, provides a kind of user credit apparatus for evaluating, and described device includes：

First acquisition module, the second behavior sequence of the first behavior sequence and at least one association user for obtaining user Row, the association user is that there are associated other users with the user in social networks；

First computing module, the first behavior sequence inputting for the user that obtains first acquisition module One Recognition with Recurrent Neural Network obtains the First ray character representation of the user；

Second behavior sequence of the second computing module, the association user for obtaining first acquisition module is defeated Enter second circulation neural network, the second sequence signature for obtaining the association user indicates；

Third computing module, the First ray character representation for obtaining first computing module and described Second sequence signature that two computing modules obtain indicates that input stack self-encoding encoder, the assemblage characteristic for obtaining the user are spread out It is raw to indicate；

4th computing module, the assemblage characteristic of the user for obtaining the third computing module, which derives, indicates defeated Enter grader, obtains the credit evaluation data of the user；

Wherein, in the first circulation neural network, the second circulation neural network and the stack self-encoding encoder Model parameter is determined after being trained using sample sequence data, the first circulation neural network and the second circulation Model parameter in neural network is identical.

The advantageous effect that technical solution provided in an embodiment of the present invention is brought is：

On the one hand, first circulation neural network and second circulation neural network are trained by using sample sequence data, made It obtains first circulation neural network and is receiving association use in the first behavior sequence or second circulation neural network for receiving user When second behavior sequence at family, first circulation neural network can extract sequence signature automatically according to the first behavior sequence, the Two Recognition with Recurrent Neural Network can extract sequence signature automatically according to the second behavior sequence, since Recognition with Recurrent Neural Network can be used in Analytical sequence data, therefore when carrying out sequence signature extraction to the first behavior sequence or the second behavior sequence, it need not be artificial Feature extraction work is participated in, and includes hidden feature by the sequence signature that Recognition with Recurrent Neural Network is drawn into, feature is covered More comprehensively, and compared with artificial extraction feature, the efficiency that feature extraction is carried out by Recognition with Recurrent Neural Network improves lid, calculates cost It reduces；On the other hand, stack self-encoding encoder is trained by using sample sequence data so that stack self-encoding encoder is followed receiving After the sequence signature of the user of ring neural network output indicates, it can indicate that automatic output combination is special according to the sequence signature of user Sign is derivative to be indicated, since stack self-encoding encoder can indicate that the automatic combination for carrying out feature derives according to the sequence signature of user, Without manually participating in work derived from feature combination, consider when avoiding engineer's combination derivative feature incomplete Problem, and improved since stack self-encoding encoder carries out efficiency derived from feature combination, so as to be suitable for large-scale spy Work derived from sign combination improves efficiency derived from feature combination, reduces calculating cost.

Description of the drawings

To describe the technical solutions in the embodiments of the present invention more clearly, make required in being described below to embodiment Attached drawing is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for For those of ordinary skill in the art, without creative efforts, other are can also be obtained according to these attached drawings Attached drawing.

Fig. 1 is the schematic diagram of the implementation environment shown in one embodiment of the invention；

Fig. 2A is the method flow diagram of user credit appraisal procedure provided by one embodiment of the present invention；

Fig. 2 B are the structural schematic diagrams of reference model provided by one embodiment of the present invention；

Fig. 3 A are the method flow diagrams for the user credit appraisal procedure that another embodiment of the present invention provides；

Fig. 3 B are the structural schematic diagrams for the reference model that another embodiment of the present invention provides；

Fig. 4 A are the flow charts of the method for trained first circulation neural network provided by one embodiment of the present invention；

Fig. 4 B are structural schematic diagrams of the first LSTM provided by one embodiment of the present invention in training；

Fig. 4 C are the structural schematic diagrams of the first LSTM units provided by one embodiment of the present invention；

Fig. 5 A are the flow charts of the method for trained stack self-encoding encoder provided by one embodiment of the present invention；

Fig. 5 B are the flow charts of the method for the training stack self-encoding encoder that another embodiment of the present invention provides；

Fig. 5 C are structural schematic diagrams of the RBM provided by one embodiment of the present invention in training；

Fig. 5 D are the schematic diagrames of RBM training flow provided by one embodiment of the present invention；

Fig. 5 E are the schematic diagrames of stack self-encoding encoder training flow provided by one embodiment of the present invention；

Fig. 6 is the block diagram of user credit apparatus for evaluating provided by one embodiment of the present invention；

Fig. 7 is the structural schematic diagram of the server provided in one embodiment of the invention.

Specific implementation mode

To make the object, technical solutions and advantages of the present invention clearer, below in conjunction with attached drawing to embodiment party of the present invention Formula is described in further detail.

Fig. 1 is the schematic diagram of the implementation environment shown in one embodiment of the invention, as shown in Figure 1, the implementation environment includes： At least two user equipmenies 110, social interaction server device 120 and data management server 130.

User equipment 110 is that the terminal of such as mobile phone, tablet computer, portable computer and desktop computer etc is set It is standby.Optionally, social class application program can be installed, optionally, social class application program supports transaction branch in user equipment 110 Pay function.3 user equipmenies 110 are schematically illustrated in Fig. 1.

After logging in social account number in the social class application program of user in the user equipment 110, user can pass through social activity Account number carries out social activity with other users, alternatively, passing through social activity in the social class application program for supporting transaction payment function Account number is traded payment activity.

Social interaction server device 120 is that the background server of social interaction server is provided for user equipment 110, and social interaction server device 120 can Can also be the server cluster being made of several servers or a cloud computing service center to be a server. Social interaction server device 120 is used to record the social record of the user equipment that each social account number logs in and each social account number.Each The social record of one of social account number or a register correspond to the behavioral data of user as the social activity account number, optional Ground, behavioral data can also be other operations that user is carried out by social class application program, such as user and association user Interactive operation etc., association user are that there are associated other users with the user in social networks.In practical applications, Mei Gehang The concrete condition of the operation behavior each time of user can be described by vector for data, includes the type (ratio of operation behavior Such as chat, transfer accounts, do shopping), the corresponding quantity of operation behavior (such as online hours, transfer amounts, shopping amount of money).

Optionally, user equipment 110 can pass through wireless network mode or cable network mode and social interaction server device 120 Establish communication connection.

Optionally, above-mentioned wireless network or cable network use standard communication techniques and/or agreement.Network be usually because Special the net, (English it may also be any network, including but not limited to LAN：Local Area Network, referred to as：LAN)、 Metropolitan Area Network (MAN) (English：Metropolitan Area Network, referred to as：MAN), wide area network (English：Wide Area Network, referred to as：WAN any combinations of), mobile, wired either wireless network, dedicated network or Virtual Private Network). In some embodiments, using including hypertext markup language (English：Hyper Text Mark-up Language, referred to as： HTML), extensible markup language (English：Extensible Markup Language, referred to as：) etc. XML technology and/or lattice Formula represents the data by network exchange.Such as security socket layer (English can additionally be used：Secure Socket Layer, referred to as：SSL), Transport Layer Security (English：Trassport Layer Security, referred to as：TLS), Virtual Private Network Network (English：Virtual Private Network, referred to as：VPN), Internet Protocol Security (English：Internet Protocol Security, referred to as：) etc. IPsec conventional encryption techniques encrypt all or some links.In further embodiments, also Customization and/or the substitution of the exclusive data communication technology can be used or supplement above-mentioned data communication technology.

Data management server 130 is the background service that the data that social interaction server device 120 is got are carried out with data processing Device, data management server 130 can be a servers, can also be the server cluster being made of several servers, or Person is a cloud computing service center.Data management server 130 is for obtaining the social account of each of storage of social interaction server device 120 The social record of number user equipment logged in and each social account number, i.e. 130 user of data management server obtain each user Behavioral data.Being established in data management server 130 has reference model, and data management server 130 is used for will be from party clothes The behavioral data for the user that business device 120 obtains inputs the reference model, obtains the credit evaluation data of the user.Data management takes The credit evaluation data for the user being calculated is sent to social interaction server device 120 by business device 130, social interaction server device 120 according to The credit evaluation data at family distinguishes limitation to the corelation behaviour of different user, such as to the lower user of credit evaluation data The behavior of application loan is negated or reduces amount.

In other exemplary embodiments, above-mentioned social interaction server device 120 and data management server 130 may be implemented same In one server or same server cluster, the present embodiment is not construed as limiting this.

Optionally, which further includes third-party server 140.Third-party server 140 is for providing each use The label data at family, label data is the label after being labeled to the credit of user, for example the label data of user is excellent (general Rate=0.6), poor (probability=0.4), usually excellent probability and the probability of difference and be 1.In practical applications, third party's service Device 140 can be the server of banking institution, or the server of third party financial institution.Optionally, third party's service Device 140 can also be the database of storage user tag data.

Data management server 130 obtains the label data of user from third-party server 140, from social interaction server device The behavioral data of 120 users obtained, trains according to the behavioral data of user and label data for calculating credit evaluation number According to reference model.

Fig. 2A is the method flow diagram of user credit appraisal procedure provided by one embodiment of the present invention, which comments Method is estimated to be illustrated in application data management server 130 shown in Fig. 1.As shown in Figure 2 A, which assesses Method may include：

Step 201, the first behavior sequence of user and the second behavior sequence of at least one association user are obtained, association is used Family is that there are associated other users with user in social networks.

The first behavior sequence of user includes continuous vector in the n sequential of user, 1 of each corresponding user of vector Behavioral data, n are positive integer.

Second behavior sequence of association user includes continuous vector in m sequential of association user, and each vector is corresponding 1 behavioral data of association user, m is positive integer.

Refer to continuously that the sequencing that vector occurs according to corresponding behavioral data is arranged in sequential.With the time Increase, behavioral data caused by user's operation is also increasing, then include in the first behavior sequence or the second behavior sequence to Amount is also increasing, i.e., the length of the first behavior sequence or the second behavior sequence can be elongated.

Step 202, by the first behavior sequence inputting first circulation neural network of user, the First ray for obtaining user is special Sign indicates.

Recognition with Recurrent Neural Network (English：Recurrent Neural Networks, referred to as：RNNs) it is used for processing sequence shape , there is self recursive connection in the hidden layer of Recognition with Recurrent Neural Network in the data of formula, the input of hidden layer not only including it is current when The output of input layer is carved, but also includes the output of last moment hidden layer.

Model parameter in first circulation neural network is determined after being trained using sample sequence data.Sample sequence Column data includes：Second sample behavior sequence of the first sample behavioral data of sample of users, at least one association sample of users With the label data of sample of users, association sample of users is that there are associated other users, marks with sample of users in social networks Label data are the labels after being labeled to the credit of sample of users, and label data is typically from bank or third party financial institution It gets.

First ray character representation is the first behavior sequence about user exported by first circulation neural network One vector of hidden feature, since First ray character representation is for indicating hidden feature, and hidden feature is by artificial Carry out the feature that can not be drawn into when foundation characteristic extraction, therefore the first behavior sequence for being drawn into of first circulation neural network Foundation characteristic is more comprehensive.

Step 203, the second behavior sequence of association user is inputted into second circulation neural network, obtains the of association user Two sequence signatures indicate.

Model parameter in second circulation neural network is identical as the model parameter in first circulation neural network, due to Model parameter in one Recognition with Recurrent Neural Network is determined after being trained using sample sequence data, first circulation neural network The function of being realized with needs with the network architecture of second circulation neural network is the same, therefore in second circulation neural network Model parameter is directly disposed as identical as the model parameter in first circulation neural network.

The expression of second sequence signature is the second behavior sequence about association user exported by second circulation neural network One vector of the hidden feature of row, since the second sequence signature is indicated for indicating that hidden feature, hidden feature are to pass through people Work carries out the feature that can not be drawn into when foundation characteristic extraction, therefore the second behavior sequence that second circulation neural network is drawn into Foundation characteristic it is more comprehensive.

Step 204, First ray character representation and the second sequence signature are indicated into input stack self-encoding encoder, obtains user Assemblage characteristic derive indicate.

First ray character representation and the expression of the second sequence signature are spliced before inputting stack self-encoding encoder, Obtain a longer vector.

Indicate to be indicated according to the sequence signature that the second behavior sequence of association user extracts due to the second sequence signature, Association user, which exists with user in social networks, to be associated with, therefore the second behavior sequence is the benefit to the first behavior sequence of user It fills, the hidden feature of some social attributes may be included.

First ray character representation and the expression of the second sequence signature input stack self-encoding encoder after being spliced, and stack is self-editing Code device can capture the interactive relation between First ray character representation and the expression of the second sequence signature, to First ray mark sheet Show that being combined feature with the expression of the second sequence signature derives.

Model parameter in stack self-encoding encoder is determined after being trained using sample sequence data.

Step 205, the assemblage characteristic of user is derived and indicates input grader, obtain the credit evaluation data of user.

Grader can be linear classifier, can also be it such as logistic regression (Logistic Regression) model The grader of his type.

The credit evaluation data of user be by grader according to the assemblage characteristic of user derive indicate obtain for pair The data that the credit of user is labeled, such as：User A=(excellent 0.8, it is poor 0.2).

Above-mentioned first circulation neural network, second circulation neural network, stack self-encoding encoder and grader are constituted to user The reference model that credit is assessed, reference model are needed using sample sequence data before structure to first circulation nerve net Network, second circulation neural network, stack self-encoding encoder, grader carry out the training of model parameter so that reference model is carrying out Accuracy and confidence level when user credit is assessed is guaranteed.In conjunction with reference to figure 2B, the present embodiment institute is schematically illustrated The structural schematic diagram for the reference model being related to.As shown in Figure 2 B, reference model 20 is followed including first circulation neural network 21, second Ring neural network 22, stack self-encoding encoder 23 and grader 24.The output of first circulation neural network 21, second circulation nerve The output of network 22 is connected with the input of stack self-encoding encoder 23, the input phase of the output and grader 24 of stack self-encoding encoder 23 Even.

In conclusion user credit appraisal procedure provided in an embodiment of the present invention, on the one hand, by using sample sequence number According to training first circulation neural network and second circulation neural network so that first circulation neural network is receiving the of user One behavior sequence or second circulation neural network are in the second behavior sequence for receiving association user, first circulation neural network Sequence signature can be extracted automatically according to the first behavior sequence, second circulation neural network can be according to the second behavior sequence certainly It is dynamic to extract sequence signature, since Recognition with Recurrent Neural Network can be used in analytical sequence data, to the first behavior sequence or When second behavior sequence carries out sequence signature extraction, feature extraction work need not be manually participated in, and by recycling nerve net The sequence signature that network is drawn into includes hidden feature, more comprehensively to the covering of feature, and compared with artificial extraction feature, by following The efficiency that ring neural network carries out feature extraction improves, and calculates cost reduction；On the other hand, it is instructed by using sample sequence data Practice stack self-encoding encoder so that stack self-encoding encoder is indicated in the sequence signature for the user for receiving Recognition with Recurrent Neural Network output Afterwards, it can indicate that automatic output assemblage characteristic derives according to the sequence signature of user to indicate, since stack self-encoding encoder being capable of root Indicate that the automatic combination for carrying out feature derives according to the sequence signature of user, without manually participating in work derived from feature combination Make, avoids and consider incomplete problem when engineer combines derivative feature, and since stack self-encoding encoder carries out feature Efficiency derived from combination improves, and so as to be suitable for work derived from large-scale feature combination, improves feature combination and spreads out Raw efficiency reduces calculating cost.

Since the first behavior sequence, the second behavior sequence can cause the length of sequence long with the growth of time, work as sequence When the length of row is long, the information transmitted in Recognition with Recurrent Neural Network can soon decay, and length memory network (English：Long Short Term Memory, referred to as：LSTM the algorithm that forgetting and intensified learning) are increased on the basis of RNNs, for length The analysis of long sequence has stronger advantage, and therefore, the present embodiment is with first circulation neural network and second circulation nerve net Network is illustrated for LSTM.LSTM includes LSTM units and average pond unit in use.

Fig. 3 A are the method flow diagram for the user credit appraisal procedure that another embodiment of the present invention provides, the user credit Appraisal procedure in application data management server 130 shown in Fig. 1 to illustrate.As shown in Figure 3A, which comments The method of estimating may include：

Step 301, the first behavior sequence of user is obtained, the first behavior sequence includes continuous vector in n sequential, often 1 behavioral data of a corresponding user of vector, n is positive integer.

Refer to continuously that the sequencing that vector occurs according to corresponding behavioral data is arranged in sequential.

It step 302, will be in n DUAL PROBLEMS OF VECTOR MAPPING to same Value space.

Due to 1 behavioral data of the corresponding user of each vector, the scale that different behavioral datas uses is different, for example goes For type, the time that behavior occurs, behavior scene be the data of different scales, the meaning of different scales is for example, false If the type of behavior is integer, the time that behavior occurs is to retain 2 significant digits, then the type of behavior and behavior occurs Time is different scales.The data of different scales generally can not be directly inputted to LSTM units, therefore for n vector, number According to management server before n vector is input to LSTM units, the value by the element of n vector is needed to be mapped to same In Value space, i.e., the value of the element of n vector is mapped as identical scale.

Step 303, it is mono- that continuous vector in n sequential after mapping is inputted to the first LSTM respectively according to temporal order Member, obtains the vector of the hidden layer output of n the first LSTM units, and the vector of n hidden layer output is formed the first matrix, In, the input of the first LSTM units of i+1 includes output and the i+1 first of the hidden layer of i-th of the oneth LSTM units The output of the input layer of LSTM units, i are positive integer, i<i+1<n.

In conjunction with reference to figure 3B, the first LSTM units 321 are a parts of the first LSTM 320, and the first LSTM 320 is in reference It is averaged pond unit 322 including the first LSTM units 321 and first during model use.Each first LSTM units 321 packet Include input layer and hidden layer.

It illustrates, it is assumed that n vector is respectively vector x according to temporal order₁, vector x₂..., vector x_n, according to sequential Sequentially, the first moment, vector x₁Input the input layer of the first LSTM units 321, the hidden layers of the first LSTM units 321 export to Measure h₁；Second moment, vector x₂The input layer of the first LSTM units 321 is inputted, the first LSTM units 321 of the second moment are hidden The input of layer is output and the first LSTM units 321 at the second moment of the hidden layer of the first LSTM units 321 at the first moment Input layer output, the hidden layer output vector h of the first LSTM units 321 of the second moment₂, behind and so on.N to After amount fully enters the first LSTM units 321, the first LSTM units 321 export the vector of n hidden layer output, are vector respectively h₁, vector h₂..., vector h_n.It follows that

Step 304, the first Input matrix first is averaged pond unit, obtains the First ray character representation of user, One average pond unit is used to be averaging every column element in the first matrix.

In conjunction with reference to figure 3B, after obtaining the first matrix, the first average pond unit 322 needs to carry out pond to the first matrix Change is handled, i.e., is averaging to every column element in the first matrix.For example,First matrix is flat by first A vector H is obtained after equal pond unit 322₁, vectorial H₁First be drawn into according to the first behavior sequence for the first LSTM320 Sequence signature indicates.

Step 305, obtain the second behavior sequence of association user, the second behavior sequence include in m sequential continuously to Amount, 1 behavioral data of each corresponding association user of vector, m is positive integer.

Association user is that there are associated other users with user in social networks.Obtain the second behavior of association user Sequence is in order to which the first behavior sequence to user supplements, and the second behavior sequence may include the implicit of some social attributes Feature.

It step 306, will be in m DUAL PROBLEMS OF VECTOR MAPPING to same Value space.

The Value space that n vector of the Value space and user that m vector of association user is mapped to is mapped to It is identical.

Step 307, continuous vector in m sequential after mapping is inputted into m the 2nd LSTM according to temporal order respectively Unit, obtains the vector of the hidden layer output of m the 2nd LSTM units, and the vector of m hidden layer output is formed the second matrix, Wherein, the input of i+1 the 2nd LSTM units includes output and the i+1 a the of the hidden layer of i-th of the 2nd LSTM units The output of the input layer of two LSTM units, i are positive integer, i<i+1<m.

In conjunction with reference to figure 3B, the 2nd LSTM units 331 are a parts of the 2nd LSTM 330, and the 2nd LSTM units 331 exist It is averaged pond unit 332 including the 2nd LSTM units 331 and second during the use of reference model.Each 2nd LSTM units 331 include input layer and hidden layer.

It illustrates, it is assumed that m vector is respectively vector x according to temporal order₁, vector x₂..., vector x_m, according to sequential Sequentially, the first moment, vector x₁Input the input layer of the 2nd LSTM units 331, the hidden layers of the 2nd LSTM units 331 export to Measure h₁；Second moment, vector x₂The input layer of the 2nd LSTM units 331 is inputted, the 2nd LSTM units 331 of the second moment are hidden The input of layer is output and the 2nd LSTM units 331 at the second moment of the hidden layer of the 2nd LSTM units 331 at the first moment Input layer output, the hidden layer output vector h of the 2nd LSTM units 331 of the second moment₂, behind and so on.M to After amount fully enters the 2nd LSTM units 331, the 2nd LSTM units 331 export the vector of m hidden layer output, are vector respectively h₁, vector h₂..., vector h_m.It follows that

Step 308, the second Input matrix second is averaged pond unit, obtains the second sequence signature table of association user Show；Wherein, when the quantity of the second behavior sequence is one, the second average pond unit is used for each column member in the second matrix Element is averaging；When the quantity of the second behavior sequence is p, the second average pond unit is used for every in each second matrix P the second sequence signatures that column element is averaging indicate to be averaging, and p is positive integer, p>1.

With user there are associated user may be one it is also likely to be multiple, therefore the second behavior sequence on social networks It is also likely to be multiple that the quantity of row, which may be one, and it is also likely to be multiple that the second obtained matrix, which may be one,.

Illustratively illustrated in conjunction with being one with reference to figure 3B, Fig. 3 B with the quantity of association user.When association user When quantity is one, the second average pond unit 332 needs to do pondization processing to the second matrix, i.e., to each column in the second matrix Element is averaging.For example,Second matrix obtains a vector after the second average pond unit 332 H₂, vectorial H₂It is indicated according to the second sequence signature that the second behavior sequence is drawn into for the 2nd LSTM 330.

When the quantity of association user is p, the second behavior sequence of p association user passes through the 2nd LSTM units respectively 331 the second matrixes of output p, the second average pond unit 332 carry out pond processing to each second matrix respectively first, obtain Then p vector carries out pond to p vector and handles to obtain a vector H₂, vectorial H₂It is the 2nd LSTM 330 according to the second row It is indicated for the second sequence signature that sequence is drawn into.

Step 309, First ray character representation and the expression of the second sequence signature are spliced, the sequence for obtaining user is special Sign indicates.

First ray character representation is a vector H of the first LSTM 320 outputs₁, the expression of the second sequence signature is second The vector H that LSTM 330 is exported₂, First ray character representation and the second sequence signature are indicated into splicing, i.e., by vectorial H₁With Vectorial H₂Splicing, obtains a longer vectorial H, and vectorial H is indicated as the sequence signature of user.

Step 310, the sequence signature of user is indicated into input stack self-encoding encoder, the i+1 layer RBM of stack self-encoding encoder Input layer input be i-th layer of RBM hidden layer output, i is positive integer, i<i+1<k.

In conjunction with reference to figure 3B, stack self-encoding encoder 340 includes k layers of limited Boltzmann machine (English：Restricted Boltzmann Machines, referred to as：RBM) 341, respectively 1RBM, 2RBM ..., kth RBM, k is positive integer, each RBM includes input layer and hidden layer.

Step 311, the output of the hidden layer of kth layer RBM is derived as the assemblage characteristic of user and is indicated.

The sequence signature of user, which derives, to be indicated after the input layer input of the 1st layer of RBM, the hidden layer of each layer of RBM it is defeated The input for going out the input layer as next layer of RBM, until kth layer RBM, i.e., the output of the hidden layer of last layer RBM obtains Vector derives as the assemblage characteristic of user to be indicated.

Step 312, the assemblage characteristic of user is derived and indicates input grader, obtain the credit evaluation data of user.

Derive as the assemblage characteristic of user in conjunction with the output with reference to figure 3B, the hidden layer of kth RBM 341 and indicates, input To grader 350, grader 350 exports the credit evaluation data of user.

In conclusion user credit appraisal procedure provided in an embodiment of the present invention, on the one hand, by using sample sequence number According to training first circulation neural network and second circulation neural network so that first circulation neural network is receiving the of user One behavior sequence or second circulation neural network are in the second behavior sequence for receiving association user, first circulation neural network Sequence signature can be extracted automatically according to the first behavior sequence, second circulation neural network can be according to the second behavior sequence certainly It is dynamic to extract sequence signature, since Recognition with Recurrent Neural Network can be used in analytical sequence data, to the first behavior sequence or When second behavior sequence carries out sequence signature extraction, feature extraction work need not be manually participated in, and by recycling nerve net The sequence signature that network is drawn into includes hidden feature, more comprehensively to the covering of feature, and compared with artificial extraction feature, by following The efficiency that ring neural network carries out feature extraction improves, and calculates cost reduction；On the other hand, it is instructed by using sample sequence data Practice stack self-encoding encoder so that stack self-encoding encoder is indicated in the sequence signature for the user for receiving Recognition with Recurrent Neural Network output Afterwards, it can indicate that automatic output assemblage characteristic derives according to the sequence signature of user to indicate, since stack self-encoding encoder being capable of root Indicate that the automatic combination for carrying out feature derives according to the sequence signature of user, without manually participating in work derived from feature combination Make, avoids and consider incomplete problem when engineer combines derivative feature, and since stack self-encoding encoder carries out feature Efficiency derived from combination improves, and so as to be suitable for work derived from large-scale feature combination, improves feature combination and spreads out Raw efficiency reduces calculating cost.

In addition, by by n DUAL PROBLEMS OF VECTOR MAPPING in the first behavior sequence to same Value space, it will be in the second behavior sequence M DUAL PROBLEMS OF VECTOR MAPPING to same Value space so that the vector in the first behavior sequence and the second behavior sequence can be input to LSTM units ensure that the normal extraction that First ray character representation and the second sequence signature indicate.

In addition, First ray character representation is extracted from the first behavior sequence by using LSTM, from the second behavior sequence The expression of the second sequence signature is extracted in row, since the input of i+1 LSTM hidden layers includes the hidden layer of i-th of LSTM unit Output and i+1 LSTM units input layer output so that the input of the hidden layer at current time is not only by current The influence of the output of the input layer at moment is also influenced by the output of the hidden layer layer of last moment, so that in the first row When long for the length of sequence or the second behavior sequence, the problem of avoiding the occurrence of unstable gradient.

In addition, by the way that the second sequence signature expression of the First ray character representation of user and association user is stitched together Sequence signature as user indicates that input stack self-encoding encoder carries out so that is indicated by the second sequence signature of association user The First ray character representation of user is supplemented, so as to get the feature in terms of social attribute, for sequence spy The extraction of sign is more comprehensive.

Reference model before the use, is needed to Recognition with Recurrent Neural Network, stack self-encoding encoder and the classification in reference model The model parameter of device is trained, and Fig. 4 A are illustrated to the training process of the first LSTM or the 2nd LSTM, and Fig. 5 A are The training process of stack self-encoding encoder is illustrated.

Fig. 4 A are the flow chart of the method for trained first circulation neural network provided by one embodiment of the present invention, this method To be illustrated in application data management server 130 shown in Fig. 1.As shown in Figure 4 A, this approach includes the following steps：

Step 401, sample sequence data are obtained, sample sequence data include：The first sample behavior sequence of sample of users, The label data of second sample behavior sequence and sample of users of at least one association sample of users, label data are used sample The credit at family be labeled after label, association sample of users is that there are other associated use in social networks and sample of users Family.

The first sample behavior sequence of sample of users includes continuous vector in n sequential of sample of users, each vector 1 behavioral data of corresponding sample of users, n is positive integer.

Second sample behavior sequence of association sample of users includes being associated with continuous vector in m sequential of sample of users, 1 behavioral data of each corresponding association sample of users of vector, m is positive integer.

Step 402, first sample behavior sequence is inputted into the first LSTM units, obtains training characteristics sequence.

First LSTM units are a parts of the first LSTM, and in conjunction with reference to figure 4B, the first LSTM 420 is in reference model structure Include that the first LSTM units 421, first are averaged the training grader 423 of pond unit 422 and first during building.Each first LSTM units 421 include input layer and hidden layer.

Training characteristics sequence is after n vector in first sample behavior sequence fully enters the first LSTM units 421 The sequence of n vector composition of the hidden layer output of one LSTM units 421.

Optionally, first sample behavior sequence needs before inputting the first LSTM units by first sample behavior sequence In sample of users n behavioral data corresponding DUAL PROBLEMS OF VECTOR MAPPING to same Value space in.

Step 403, training characteristics sequence inputting first is averaged pond unit, obtains the expression of training sample sequence signature.

First average pond unit is used to carry out pond processing to n vector in training characteristics sequence, i.e., special to training The n vector levied in sequence is averaging, and the vector obtained after averaging is indicated as training sample sequence signature.

Step 404, training sample sequence signature is indicated into input the first training grader, obtains the first prediction data.

In the building process of reference model, the parameter needs in the first LSTM are trained, the mounting of the first tail portions LSTM First training grader, for according to training sample sequence signature indicate output the first prediction data, then the first LSTM according to The parameter in the first LSTM units of difference pair between first prediction data and the label data of sample of users is adjusted.First The parameter of LSTM uses random initializtion in the initial state, and according to the parameter of random initializtion, the first LSTM outputs first are pre- Measured data, first training grader according to the parameter in the first LSTM units of the first prediction data and the difference pair of label data into Row adjustment.

In conjunction with reference to figure 4C, it illustrates the structure of the first LSTM units, the first LSTM units 421 include mnemon 431, input gate 432, out gate 433, forget that door 434 connects 435 with self recurrence.Input gate 432 allows to input mnemon 431 signal changes the signal of mnemon 431 or shielding input mnemon 431, and out gate 433 allows mnemon 431 Output the lower mnemon of effect of signals or screen memory unit 431 output signal, forget door 434 for manage memory singly The situation of self connection of member 431 whenever necessary, is forgetting door 434 for controlling the shape before mnemon 431 forgets itself State.

Optionally, the first training grader for being trained to the first LSTM is usually SoftMax (soft maximum).

Step 405, the first prediction data and label data are substituted into first-loss function, whether judges first-loss function Converge to minimum.

First prediction data and the difference of label data are measured by cross entropy, according to cross entropy and penalty term composition first Loss function.Training objective to the first LSTM units is first-loss function convergence to minimum.

Step 406, when first-loss function does not converge to minimum, error backpropagation algorithm pair first is utilized The parameter of LSTM units is adjusted, until first-loss function convergence to minimum.

First-loss function does not converge to minimum, shows that the difference of the first prediction data and label data is excessive, then Show that the parameter of current first LSTM units is improper, the parameter needs of the first LSTM units are adjusted, and adjustment target is to make First-loss function convergence is to minimum.

Step 407, when first-loss function convergence is to minimum, the parameter of the first LSTM units after adjustment is determined For the model parameter of first circulation neural network.

First-loss function convergence to minimum, show the parameter of the first LSTM units after adjustment be it is suitable, therefore The parameter of the first LSTM units after adjustment is determined as first circulation neural network or the model parameter of the first LSTM.

In conclusion the method for trained first circulation neural network provided in an embodiment of the present invention, by by sample sequence Data are obtained by the first LSTM units, the first average pond unit and the first training grader in first circulation neural network First prediction data, according between the first prediction data and the label data of sample of users the ginseng of the first LSTM units of difference pair Number is adjusted, until sample sequence data pass through the first LSTM units in first circulation neural network, the first average pond The difference of the label data for the first prediction data and sample of users that unit and the first training grader obtain is less than first threshold When, show that the parameter of the first LSTM units is trained to model parameter, so as to the first circulation nerve net that will be trained Network is determined as the model for carrying out sequence signature extraction to the behavior sequence of user.

It should be added that due to the model parameter phase of second circulation neural network and first circulation neural network Together, when first circulation neural network and second circulation neural network are LSTM, the first LSTM units and the 2nd LSTM units Parameter is identical, is trained and is completed by the step shown in Fig. 4 A due to the parameter of the first LSTM units, optionally, the 2nd LSTM units It need not be trained, the parameters of the 2nd LSTM units is directly disposed as identical as the parameter after the first LSTM module trainings.

It is finished in the first LSTM and the 2nd LSTM training, after obtaining the model parameter of the first LSTM and the 2nd LSTM, training Input of the output of successful first LSTM and the 2nd LSTM as stack self-encoding encoder, stack self-encoding encoder proceed by instruction Practice.

Fig. 5 A are the flow charts of the method for trained stack self-encoding encoder provided by one embodiment of the present invention, and this method is to answer It is illustrated in data management server 130 shown in FIG. 1.As shown in Figure 5A, this method may comprise steps of：

Step 501, first sample behavior sequence is inputted into the first LSTM units, obtains the fisrt feature sequence of sample of users Row.

Optionally, for first sample behavior sequence before inputting LSTM units, needing will be in first sample behavior sequence In the n behavioral data corresponding DUAL PROBLEMS OF VECTOR MAPPING to same Value space of sample of users.

Step 502, fisrt feature sequence inputting first is averaged pond unit, obtains the expression of first sample sequence signature.

Step 503, the second sample behavior sequence is inputted into the 2nd LSTM units, obtains the second feature of association sample of users Sequence.

Optionally, the second sample behavior sequence needs before inputting the 2nd LSTM units by the second sample behavior sequence In association sample of users m behavioral data corresponding DUAL PROBLEMS OF VECTOR MAPPING to same Value space in.

Step 504, second feature sequence inputting second is averaged pond unit, obtains the second sample sequence character representation.

Optionally, when the quantity of the second sample behavior sequence is one, the second average pond unit is used for the second spy Sequence is levied to be averaging；When the quantity of the second sample behavior sequence is p, the second average pond unit is used for special to each second P the second sample sequence character representations that sign sequence is averaging are averaging, and p is positive integer, p>1.

Step 505, first sample sequence signature is indicated and the second sample sequence character representation splices, obtain sample The sequence signature of user indicates.

Step 501 is similar to step 309 with step 301 to step 505, the first behavior sequence in step 301 to step 309 Row and the second behavior sequence are respectively replaced by the first sample behavior sequence in step 501 to step 505 and the second sample row For sequence, the specific implementation indicated for determining the sequence signature of sample of users just repeats no more here.

Step 506, the sequence signature for mixing the sample with family indicates input stack self-encoding encoder, the sequence signature table of sample of users Show and is calculated according to sample sequence data.

The calculating that the sequence signature of sample of users indicates refers to step 501 to step 505.In the first LSTM and second After LSTM is trained, the sequence signature expression of sample of users need not be input to the first training grader, be directly inputted to stack Formula self-encoding encoder.

The training of stack self-encoding encoder includes two stages, and the first stage is the pre-training stage, refers to step 507；The Two-stage is the accurate adjustment stage, refers to step 508.

Step 507, in the pre-training stage, unsupervised learning training is carried out to every layer of RBM respectively, after obtaining pre-training RBM parameters.

In the pre-training stage, every layer of RBM is individually trained, and after the completion of preceding layer RBM is trained, and later layer RBM is Proceed by training.

RBM parameters include the weight matrix and bias vector of the hidden layer of RBM.

Optionally, step 507 can be replaced by step as shown in Figure 5 B：

Predetermined characteristic is indicated the input layer of i-th layer of RBM of input, root by step 507a when being trained to i-th layer of RBM According to the data of the output layer of i-th layer of RBM of the i-th weight matrix and the calculating of the i-th bias vector of i-th layer of RBM.

Training objective for each layer of RBM is that the data of input layer are approximately equal to the data of output layer.In conjunction with reference chart 5C, Fig. 5 C show that structures of the RBM in training, every layer of RBM 520 include input layer 521, hidden layer 522 and output layer 523. In conjunction with the training process schematic diagram for showing RBM with reference to figure 5D, Fig. 5 D, the training of RBM 520 includes two ranks of coding and decoding The vector x that input layer 521 inputs is mapped as the vectorial h of hidden layer 522 by section, encoder 531, and decoder 532 is by hidden layer 522 Vectorial h be mapped as the vector x of output layer_rec。

Training objective to every layer of RBM 520 is x ≈ x_rec, for x and x_recBetween difference by cross entropy L (x, x_rec) measure, cross entropyWherein, d_xIndicate vector Including dimension, x_(i)Indicate vector x in the data of i-th dimension degree, x_rec(i)Indicate vector x_recIn the data of i-th dimension degree.Lose letter Number Loss (θ) is made of cross entropy and penalty term, loss functionWherein, D_nTable Show that vector space, λ indicate that the coefficient of penalty term, Penalty indicate penalty term, penalty term is configured according to demand, is artificial Preset formula.

After the completion of each layer of RBM 520 is trained, output layer 523 is abandoned, by exporting to next layer for hidden layer 522 The input layer 521 of RBM 520.

Optionally, input of the output of the hidden layer of i-th layer of RBM as the input layer of i+1 layer RBM, i is positive integer, i Predetermined characteristic expression is that the sequence signature of sample of users indicates when being 1, and predetermined characteristic expression is (i-1)-th layer of RBM when i is more than 1 The output of hidden layer, i<i+1<k.

The data of the input layer of i-th layer of RBM and the data of output layer are substituted into the second loss function, judged by step 507b Whether the second loss function converges to minimum.

Data counting loss function according to the data of input layer and output layer be in order to calculate the data of input layer with it is defeated Go out the difference between the data of layer.

Step 507c, when the second loss function does not converge to minimum, adjustment the i-th weight matrix and i-th is biased towards The data of the output layer of i-th layer of RBM after the data of the input layer of i-th layer of RBM and adjustment are substituted into the second loss function by amount, Until the second loss function converges to minimum.

Step 507d biases the i-th weight matrix and i-th after adjustment when the second loss function converges to minimum Vector is determined as the RBM parameters after pre-training.

After the completion of the parameter training of i-th layer of RBM, remove the output layer of i-th layer of RBM, by the hidden layer of i-th layer of RBM It exports to the input layer of i+1 layer RBM, i+1 layer RBM is trained using identical method.

Step 508, in the accurate adjustment stage, supervised learning training is carried out to k layers of RBM in conjunction with the second training grader, there is prison Learning training is superintended and directed for adjusting the RBM parameters after pre-training.

Due in the pre-training stage, the weight matrix and bias vector of every layer of RBM carried out adjustment, therefore in accurate adjustment rank Section, stack self-encoding encoder carry out more accurate adjustment according to label data to every layer of RBM parameter.

Optionally, step 508 can be replaced by step as shown in Figure 5 B：

The output of the hidden layer of kth layer RBM is derived as training assemblage characteristic and indicates to input to the second instruction by step 508a Practice grader, obtains the second prediction data.

In the accurate adjustment stage, stack self-encoding encoder tail portion needs mounting the second training grader, the instruction for being exported according to RBM Practice assemblage characteristic and derive expression the second prediction data of output, then stack self-encoding encoder is according to the second prediction data and sample of users Label data between difference every layer of RBM parameter is further adjusted.

Optionally, the second training grader for being trained to stack self-encoding encoder is typically SoftMax.

Second prediction data and label data are substituted into third loss function, judge that third loss function is by step 508b It is no to converge to minimum.

Second prediction data and the difference of label data are measured by cross entropy, and third is formed according to cross entropy and penalty term Loss function.Training objective to stack self-encoding encoder is that third loss function converges to minimum.

Step 508c, when third loss function does not converge to minimum, using error backpropagation algorithm to every layer The weight matrix and bias vector of the hidden layer of RBM are adjusted, until third loss function converges to minimum.

Third loss function does not converge to minimum, shows that the difference of the second prediction data and label data is excessive, then Show that current each layer RBM parameters are improper, every layer of RBM parameters needs are adjusted, and adjustment target is the convergence of third loss function To minimum.

Step 508d, when third loss function converges to minimum, by the weight of the hidden layer of every layer of RBM after adjustment Matrix and bias vector are determined as the model parameter of stack self-encoding encoder.

Third loss function converges to minimum, shows that each layer RBM parameters after adjustment are suitable, therefore will be after adjustment Every layer of RBM parameter be determined as the model parameter of stack self-encoding encoder.

According to the training method of stack self-encoding encoder described in the present embodiment, Fig. 5 E show the instruction of stack self-encoding encoder Practice process schematic.As shown in fig. 5e, in the pre-training stage 50, k layers of RBM 520 are trained respectively, and RBM 520 is only in Fig. 5 E Show input layer 521 and hidden layer 522, after the completion of preceding layer RBM 520 is trained, preceding layer RBM 520 is by hidden layer 522 It exports to the input layer 521 of next layer of RBM 520.After the pre-training stage 50 is to k layers of RBM 520 all training, accurate adjustment rank The hidden layer 522 of every layer of RBM 520 is extracted the training for carrying out the accurate adjustment stage 51 by section 51, into hiding for accurate adjustment stage 51 The parameter of layer 522 is obtained by the training of pre-training stage 50.The accurate adjustment stage 51 mounts behind kth layer hidden layer 522 Two training graders 530.

Optionally, after the completion of stack self-encoding encoder is trained, the grader in reference model can be that the second training is classified Device, or other kinds of grader.

After the completion of training, the assemblage characteristic of stack self-encoding encoder output derives to be indicated to be input to classification stack self-encoding encoder Device carries out the training of reference model for grader according to label data.

In conclusion the method for trained stack self-encoding encoder provided in an embodiment of the present invention, by first circulation nerve After network and second circulation neural metwork training, first sample behavior sequence is inputted into first circulation neural network, is obtained First sample sequence signature indicates, the second sample behavior sequence is inputted second circulation neural network, obtains the second sample sequence Character representation, by the sequence for the sample of users that first sample sequence signature indicates and the second sample sequence character representation splices Input of the character representation as stack self-encoding encoder, stack self-encoding encoder exports the second prediction data in training, according to second The label data of prediction data and sample of users is adjusted the parameter of every layer of RBM in stack self-encoding encoder, until sample is used The sequence signature at family indicates the difference of the label data of the second prediction data and sample of users that are obtained by stack self-encoding encoder When less than second threshold, show that the parameter of every layer of RBM is trained to model parameter, it is self-editing so as to the stack that will train Code device is determined as indicating that being combined feature derives the model indicated for the sequence signature of the user to extraction.

Fig. 6 is the block diagram of user credit apparatus for evaluating provided by one embodiment of the present invention, which comments Device is estimated to be illustrated in application data management server 130 shown in Fig. 1.As shown in fig. 6, user credit assessment dress Set including：

First acquisition module 610, for realizing above-mentioned steps 201, step 301, step 305 and it is any other implicit or The disclosed and relevant function of acquisition.

First computing module 620, for realizing above-mentioned steps 202 and any other implicit or disclosed related to calculating Function.

Second computing module 630, for realizing above-mentioned steps 203 and any other implicit or disclosed related to calculating Function.

Third computing module 640, for realizing above-mentioned steps 204 and any other implicit or disclosed related to calculating Function.

4th computing module 650, for realizing above-mentioned steps 205, step 312 and it is any other implicit or it is disclosed with Calculate relevant function.

Optionally, the first computing module 620 includes：First computing unit and the second computing unit.

First computing unit, for realizing above-mentioned steps 303 and any other implicit or disclosed relevant with calculating Function.

Second computing unit, for realizing above-mentioned steps 304 and any other implicit or disclosed relevant with calculating Function.

Optionally, the second computing module 630 includes：Third computing unit and the 4th computing unit.

Third computing unit, for realizing above-mentioned steps 307 and any other implicit or disclosed relevant with calculating Function.

4th computing unit, for realizing above-mentioned steps 308 and any other implicit or disclosed relevant with calculating Function.

Optionally, which further includes：Second acquisition module, the 5th computing module, the 6th computing module, the 7th calculate mould Block, judgment module, adjustment module, determining module.

Second acquisition module, for realizing above-mentioned steps 401 and any other implicit or disclosed relevant with acquisition Function.

5th computing module, for realizing above-mentioned steps 402 and any other implicit or disclosed relevant with calculating Function.

6th computing module, for realizing above-mentioned steps 403 and any other implicit or disclosed relevant with calculating Function.

7th computing module, for realizing above-mentioned steps 404 and any other implicit or disclosed relevant with calculating Function.

Judgment module, for realizing above-mentioned steps 405 and it is any other implicit or it is disclosed with judge relevant function.

Module is adjusted, for realizing above-mentioned steps 406 and any other implicit or disclosed and relevant function of adjustment.

Determining module, for realizing above-mentioned steps 407 and any other implicit or disclosed and relevant function of adjustment.

Optionally, third computing module 640, including concatenation unit, input unit, the first determination unit.

Concatenation unit, for realizing above-mentioned steps 309 and any other implicit or disclosed and relevant function of splicing.

Input unit, for realizing above-mentioned steps 310 and any other implicit or disclosed and relevant function of input.

First determination unit, for realizing above-mentioned steps 311 and it is any other implicit or it is disclosed with determine it is relevant Function.

Optionally, which further includes：Input module, the first training module, the second training module.

Input module, for realizing above-mentioned steps 506 and any other implicit or disclosed and relevant function of input.

First training module, for realizing above-mentioned steps 507 and it is any other implicit or it is disclosed with train it is relevant Function.

Second training module, for realizing above-mentioned steps 508 and it is any other implicit or it is disclosed with train it is relevant Function.

Optionally, the first training module includes：5th computing unit, the first judging unit, the first adjustment unit, second are really Order member.

5th computing unit, for realizing above-mentioned steps 507a and any other implicit or disclosed relevant with calculating Function.

First judging unit, for realizing above-mentioned steps 507b and it is any other implicit or it is disclosed with judge it is relevant Function.

The first adjustment unit, for realizing above-mentioned steps 507c and any other implicit or disclosed relevant with adjustment Function.

Second determination unit, for realizing above-mentioned steps 507d and it is any other implicit or it is disclosed with determine it is relevant Function.

Optionally, the second training module includes：6th computing unit, second judgment unit, second adjustment unit, third are true Order member.

6th computing unit, for realizing above-mentioned steps 508a and any other implicit or disclosed relevant with calculating Function.

Second judgment unit, for realizing above-mentioned steps 508b and it is any other implicit or it is disclosed with judge it is relevant Function.

Second adjustment unit, for realizing above-mentioned steps 508c and any other implicit or disclosed relevant with adjustment Function.

Third determination unit, for realizing above-mentioned steps 508d and it is any other implicit or it is disclosed with determine it is relevant Function.

Optionally, which further includes：8th computing module, the 9th computing module, the tenth computing module, the 11st calculate Module, concatenation module.

8th computing module, for realizing above-mentioned steps 501 and any other implicit or disclosed relevant with calculating Function.

9th computing module, for realizing above-mentioned steps 502 and any other implicit or disclosed relevant with calculating Function.

Tenth computing module, for realizing above-mentioned steps 503 and any other implicit or disclosed relevant with calculating Function.

11st computing module, for realizing above-mentioned steps 504 and any other implicit or disclosed related to calculating Function.

Concatenation module, for realizing above-mentioned steps 505 and any other implicit or disclosed and relevant function of splicing.

Optionally, the first acquisition module 610, including：First map unit and the second map unit.

First map unit, for realizing above-mentioned steps 302 and any other implicit or disclosed relevant with mapping Function.

Second map unit, for realizing above-mentioned steps 306 and any other implicit or disclosed relevant with mapping Function.

In conclusion user credit apparatus for evaluating provided in an embodiment of the present invention, is trained by using sample sequence data The model parameter of first circulation neural network, second circulation neural network and stack self-encoding encoder, obtains for user credit The first behavior sequence inputting first circulation neural network of user is obtained First ray mark sheet by the reference model assessed Show, the second behavior sequence input second circulation neural network of association user, which is obtained the second sequence signature, to be indicated, by the first sequence Row character representation and the second sequence signature indicate that input stack self-encoding encoder obtains assemblage characteristic and derives expression, and assemblage characteristic is spread out It is raw to indicate that input grader obtains the credit evaluation data of user.Due to the first behavior sequence of user and association user When second behavior sequence carries out sequence signature extraction and assemblage characteristic derivative, need not manually it participate in, but certainly by computer It is dynamic to realize, the considerations of derivative feature causes is combined not by artificial extraction feature and engineer in the prior art to solve Comprehensively, efficiency is low, calculates problem of high cost, has reached automatic carry out sequence signature and has extracted and feature combination derivative so that is special Sign covers comprehensive effect, while improving efficiency, reduces calculating cost.

It should be noted that：The user credit apparatus for evaluating provided in above-described embodiment when assessing user credit, only with The division progress of above-mentioned each function module, can be as needed and by above-mentioned function distribution by not for example, in practical application Same function module is completed, i.e., the internal structure of server is divided into different function modules, described above complete to complete Portion or partial function.In addition, user credit apparatus for evaluating and user credit appraisal procedure embodiment that above-described embodiment provides Belong to same design, specific implementation process refers to embodiment of the method, and which is not described herein again.

Fig. 7 is the structural schematic diagram of the server provided in one embodiment of the invention.The server can be shown in Fig. 1 Data management server 130, can also be social interaction server device 120, can also be third-party server 140.Specifically：Clothes Business device 700 includes central processing unit (CPU) 701 including random access memory (RAM) 702 and read-only memory (ROM) 703 system storage 704, and connect the system bus 705 of system storage 704 and central processing unit 701.The clothes Business device 700 further includes the basic input/output (I/O systems) of transmission information between each device helped in computer 706, and for the mass-memory unit 707 of storage program area 713, application program 714 and other program modules 715.

The basic input/output 706 includes display 708 for showing information and inputs letter for user The input equipment 709 of such as mouse, keyboard etc of breath.The wherein described display 708 and input equipment 709 are all by being connected to The i/o controller 710 of system bus 705 is connected to central processing unit 701.The basic input/output 706 Can also include input and output controller 710 for receive and handle from keyboard, mouse or electronic touch pen etc. it is multiple its The input of his equipment.Similarly, i/o controller 710 also provides output to display screen, printer or other kinds of defeated Go out equipment.

The mass-memory unit 707 is by being connected to the bulk memory controller (not shown) of system bus 705 It is connected to central processing unit 701.The mass-memory unit 707 and its associated computer-readable medium are server 700 provide non-volatile memories.That is, the mass-memory unit 707 may include such as hard disk or CD-ROM The computer-readable medium (not shown) of driver etc.

Without loss of generality, the computer-readable medium may include computer storage media and communication media.Computer Storage medium includes information such as computer-readable instruction, data structure, program module or other data for storage The volatile and non-volatile of any method or technique realization, removable and irremovable medium.Computer storage media includes RAM, ROM, EPROM, EEPROM, flash memory or other solid-state storages its technologies, CD-ROM, DVD or other optical storages, tape Box, tape, disk storage or other magnetic storage apparatus.Certainly, skilled person will appreciate that the computer storage media It is not limited to above-mentioned several.Above-mentioned system storage 704 and mass-memory unit 707 may be collectively referred to as memory.

According to various embodiments of the present invention, the server 700 can also be arrived by network connections such as internets Remote computer operation on network.Namely server 700 can be by the network interface that is connected on the system bus 705 Unit 711 is connected to network 712, in other words, can also be connected to using Network Interface Unit 711 other kinds of network or Remote computer system (not shown).

The embodiment of the present invention additionally provides a kind of computer readable storage medium, which can be Computer readable storage medium included in memory in above-described embodiment；Can also be individualism, eventually without supplying Computer readable storage medium in end.There are one the computer-readable recording medium storages or more than one program, this one A either more than one program is used for executing above-mentioned user credit appraisal procedure by one or more than one processor.

The embodiments of the present invention are for illustration only, can not represent the quality of embodiment.

One of ordinary skill in the art will appreciate that realizing that all or part of step of above-described embodiment can pass through hardware It completes, relevant hardware can also be instructed to complete by program, the program can be stored in a kind of computer-readable In storage medium, storage medium mentioned above can be read-only memory, disk or CD etc..

The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all the present invention spirit and Within principle, any modification, equivalent replacement, improvement and so on should all be included in the protection scope of the present invention.

Claims (20)

1. a kind of user credit appraisal procedure, which is characterized in that the method includes：

The second behavior sequence of the first behavior sequence and at least one association user of user is obtained, the association user is in society Handing over network, there are associated other users with the user；

By the first behavior sequence inputting first circulation neural network of the user, the First ray mark sheet of the user is obtained Show；

Second behavior sequence of the association user is inputted into second circulation neural network, obtains the second sequence of the association user Row character representation；

The First ray character representation and second sequence signature are indicated into input stack self-encoding encoder, obtain the user Assemblage characteristic derive indicate；

The assemblage characteristic of the user is derived and indicates input grader, obtains the credit evaluation data of the user；

Wherein, the model in the first circulation neural network, the second circulation neural network and the stack self-encoding encoder Parameter is determined after being trained using sample sequence data, the first circulation neural network and second circulation nerve Model parameter in network is identical.

2. according to the method described in claim 1, it is characterized in that, the first circulation neural network is remembered including the first length Network LSTM units and the first average pond unit, each first LSTM units include input layer and hidden layer；Described One behavior sequence includes continuous vector in n sequential, and 1 behavioral data of each corresponding user of vector, n is just Integer；

The first behavior sequence inputting first circulation neural network by the user, the First ray for obtaining the user are special Sign expression, including：

Continuous vector in the n sequential is inputted into n the first LSTM units according to temporal order respectively, obtains n The vector of the n hidden layer outputs is formed the first matrix by the vector of the hidden layer output of the first LSTM units, In, the input of i+1 the first LSTM units include the hidden layer of i-th of the oneth LSTM unit output and i-th+ The output of the input layer of 1 the first LSTM unit, i are positive integer, i<i+1<n；

By the described in first Input matrix first average pond unit, the First ray character representation of the user, institute are obtained The first average pond unit is stated to be used to be averaging every column element in first matrix.

3. according to the method described in claim 1, it is characterized in that, the second circulation neural network includes the 2nd LSTM units With the second average pond unit, each 2nd LSTM units include input layer and hidden layer；The second behavior sequence packet Continuous vector in m sequential is included, 1 behavioral data of each corresponding association user of vector, m is positive integer；

Second behavior sequence by the association user inputs second circulation neural network, obtains the of the association user The expression of two sequence signatures, including：

Continuous vector in the m sequential is inputted into m the 2nd LSTM units according to temporal order respectively, obtains m The vector of the m hidden layer outputs is formed the second matrix by the vector of the hidden layer output of the 2nd LSTM units, In, the input of i+1 the 2nd LSTM units include the hidden layer of i-th of the 2nd LSTM unit output and i-th+ The output of the input layer of 1 the 2nd LSTM unit, i are positive integer, i<i+1<m；

By the described in second Input matrix second average pond unit, the second sequence signature table of the association user is obtained Show；Wherein, when the quantity of second behavior sequence is one, the described second average pond unit is used for second square Every column element in battle array is averaging；When the quantity of second behavior sequence is p, the described second average pond unit is used for P second sequence signatures that every column element in each second matrix is averaging are indicated to be averaging, p is just Integer, p>1.

4. method according to any one of claims 1 to 3, which is characterized in that the first circulation neural network is in training Including the first LSTM units, the first average pond unit and the first training grader；

The training process of the first circulation neural network includes the following steps：

The sample sequence data are obtained, the sample sequence data include：The first sample behavior sequence of sample of users, at least The label data of second sample behavior sequence and the sample of users of one association sample of users, the label data is to institute State sample of users credit be labeled after label, the association sample of users is deposited in social networks and the sample of users In associated other users；

The first sample behavior sequence is inputted into the first LSTM units, obtains training characteristics sequence；

By the described in the training characteristics sequence inputting first average pond unit, the expression of training sample sequence signature is obtained；

The training sample sequence signature is indicated into input the first training grader, obtains the first prediction data；

First prediction data and the label data are substituted into first-loss function, whether judge the first-loss function Converge to minimum；

It is mono- to the first LSTM using error backpropagation algorithm when the first-loss function does not converge to minimum The parameter of member is adjusted, until the first-loss function convergence to minimum；

When the first-loss function convergence is to minimum, the parameter of the first LSTM units after adjustment is determined as institute State the model parameter of first circulation neural network.

5. according to the method described in claim 1, it is characterized in that, the stack self-encoding encoder includes k layers of limited Boltzmann Machine RBM, each RBM includes input layer and hidden layer, and k is positive integer；

It is described that the First ray character representation and second sequence signature are indicated into input stack self-encoding encoder, it obtains described The assemblage characteristic of user, which derives, to be indicated, including：

The First ray character representation and second sequence signature expression are spliced, the sequence for obtaining the user is special Sign indicates；

The sequence signature of the user is indicated to input the stack self-encoding encoder, the i+1 layer institute of the stack self-encoding encoder The output for the hidden layer that the input for stating the input layer of RBM is i-th layer of RBM, i is positive integer, i<i+1<k；

The output of the hidden layer of the kth layer RBM is derived as the assemblage characteristic of the user and is indicated.

6. method according to claim 1 or 5, which is characterized in that the stack self-encoding encoder training when include k layers by The training graders of Boltzmann machine RBM and second are limited, k is positive integer；

The training process of the stack self-encoding encoder includes the following steps：

The sequence signature for mixing the sample with family indicates that the input stack self-encoding encoder, the sequence signature expression of the sample of users are It is calculated according to the sample sequence data；

In the pre-training stage, unsupervised learning training is carried out to RBM described in every layer respectively, obtains the RBM parameters after pre-training；

In the accurate adjustment stage, supervised learning training is carried out in conjunction with the second training grader RBM described to k layers, it is described to have prison Learning training is superintended and directed for adjusting the RBM parameters after the pre-training.

7. according to the method described in claim 6, it is characterized in that, each RBM includes input layer, hidden layer in training And output layer；

It is described that unsupervised learning training is carried out to RBM described in every layer respectively in the pre-training stage, obtain the ginsengs of the RBM after pre-training Number, including：

When being trained to i-th layer of RBM, predetermined characteristic is indicated to the input layer of i-th layer of RBM of input, according to i-th The i-th weight matrix and the i-th bias vector of the layer RBM calculate the data of the output layer of i-th layer of RBM；

The data of the input layer of i-th layer of RBM and the data of output layer are substituted into the second loss function, judge second damage Lose whether function converges to minimum；

When second loss function does not converge to minimum, adjusts i-th weight matrix and described i-th and be biased towards The data of the output layer of i-th layer of RBM after the data of the input layer of i-th layer of RBM and adjustment are substituted into described the by amount Two loss functions, until second loss function converges to minimum；

When second loss function converges to minimum, i-th weight matrix and described i-th after adjustment is biased towards Amount is determined as the RBM parameters after the pre-training；

Wherein, input of the output of the hidden layer of i-th layer of RBM as the input layer of the i+1 layer RBM, i are just whole Number, the predetermined characteristic indicates it is that the sequence signature of the sample of users indicates that i is more than predetermined characteristic table when 1 when i is 1 Show the output for the hidden layer for being (i-1)-th layer of RBM, i<i+1<k.

8. according to the method described in claim 6, it is characterized in that, described in the accurate adjustment stage, in conjunction with the second training classification Device RBM described to k layers carries out supervised learning training, including：

The output of the hidden layer of the kth layer RBM is derived as training assemblage characteristic and indicates to input to second training point Class device obtains the second prediction data；

Second prediction data and the label data are substituted into third loss function, whether judge the third loss function Converge to minimum；

When the third loss function does not converge to minimum, using error backpropagation algorithm to RBM described in every layer The weight matrix and bias vector of hidden layer are adjusted, until the third loss function converges to minimum；

When the third loss function converges to minimum, by the weight matrix of the hidden layer of every layer of RBM after adjustment It is determined as the model parameter of the stack self-encoding encoder with bias vector.

9. according to the method described in claim 6, it is characterized in that, the first circulation neural network includes the first LSTM units With the first average pond unit；The second circulation neural network includes the 2nd LSTM units and the second average pond unit；Institute State the second sample of first sample behavior sequence, at least one association sample of users that sample sequence data include the sample of users The label data of this behavior sequence and the sample of users, the label data are labeled to the credit of the sample of users Label afterwards, the association sample of users are that there are associated other users with the sample of users in social networks；

The sequence signature for mixing the sample with family indicates to input before the stack self-encoding encoder, further includes：

The first sample behavior sequence is inputted into the first LSTM units, obtains the fisrt feature sequence of the sample of users Row；

By the described in the fisrt feature sequence inputting first average pond unit, the expression of first sample sequence signature is obtained；

The second sample behavior sequence is inputted into the 2nd LSTM units, obtains the second feature of the association sample of users The parameter of sequence, the 2nd LSTM units is identical as the parameter of the first LSTM units；

By the described in the second feature sequence inputting second average pond unit, the second sample sequence character representation is obtained；

The first sample sequence signature is indicated and the second sample sequence character representation splices, obtains the sample The sequence signature of user indicates that the sequence signature expression of the sample of users is the input of the stack self-encoding encoder；

Wherein, when the quantity of the second sample behavior sequence is one, the described second average pond unit is used for described Second feature sequence is averaging；When the quantity of the second sample behavior sequence is p, the described second average pond unit is used It is averaging in the p being averaging to each second feature sequence the second sample sequence character representations, p is just whole Number, p>1.

10. method according to any one of claims 1 to 9, which is characterized in that it is described obtain user the first behavior sequence and Second behavior sequence of at least one association user, including：

N vector included by the first behavior sequence of the user is obtained, by the n DUAL PROBLEMS OF VECTOR MAPPING to same Value space In, n is positive integer；

M included by the second behavior sequence of association user vector is obtained, by the m DUAL PROBLEMS OF VECTOR MAPPING to described same In Value space, m is positive integer.

11. a kind of user credit apparatus for evaluating, which is characterized in that described device includes：

First acquisition module, the second behavior sequence of the first behavior sequence and at least one association user for obtaining user, The association user is that there are associated other users with the user in social networks；

First behavior sequence inputting first of the first computing module, the user for obtaining first acquisition module follows Ring neural network obtains the First ray character representation of the user；

Second computing module, the second behavior sequence of the association user for obtaining first acquisition module input the Two Recognition with Recurrent Neural Network, the second sequence signature for obtaining the association user indicate；

Third computing module, based on the First ray character representation and described second by obtaining first computing module It calculates second sequence signature that module obtains and indicates input stack self-encoding encoder, obtain the assemblage characteristic derivative table of the user Show；

4th computing module, the assemblage characteristic of the user for obtaining the third computing module, which derives, indicates input point Class device obtains the credit evaluation data of the user；

Wherein, the model in the first circulation neural network, the second circulation neural network and the stack self-encoding encoder Parameter is determined after being trained using sample sequence data, the first circulation neural network and second circulation nerve Model parameter in network is identical.

12. according to the devices described in claim 11, which is characterized in that the first circulation neural network is remembered including the first length Recall network LSTM units and the first average pond unit, each first LSTM units include input layer and hidden layer；It is described First behavior sequence includes continuous vector in n sequential, 1 behavioral data of each corresponding user of vector, and n is Positive integer；

First computing module, including：

First computing unit, for continuous vector in the n sequential to be inputted n described first according to temporal order respectively LSTM units obtain the vector of the hidden layer output of n the first LSTM units, by the vector of the n hidden layer output Form the first matrix, wherein the input of i+1 the first LSTM units includes the hidden of i-th the oneth LSTM unit The output of the output of layer and the input layer of i+1 the first LSTM units is hidden, i is positive integer, i<i+1<n；

Second computing unit, the first average pond described in first Input matrix for obtaining first computing unit Unit, obtains the First ray character representation of the user, and the described first average pond unit is used for in first matrix Every column element be averaging.

13. according to the devices described in claim 11, which is characterized in that the second circulation neural network includes the 2nd LSTM mono- Member and the second average pond unit, each 2nd LSTM units include input layer and hidden layer；Second behavior sequence Including continuous vector in m sequential, 1 behavioral data of each corresponding association user of vector, m is positive integer；

Second computing module, including：

Third computing unit, for continuous vector in the m sequential to be inputted m described second according to temporal order respectively LSTM units obtain the vector of the hidden layer output of m the 2nd LSTM units, by the vector of the m hidden layer output Form the second matrix, wherein the input of i+1 the 2nd LSTM units includes the hidden of i-th the 2nd LSTM unit The output of the output of layer and the input layer of i+1 the 2nd LSTM units is hidden, i is positive integer, i<i+1<m；

4th computing unit, the second average pond described in second Input matrix for obtaining the third computing unit Unit, the second sequence signature for obtaining the association user indicate；Wherein, when the quantity of second behavior sequence is one When, the described second average pond unit is used to be averaging every column element in second matrix；When the second behavior sequence When the quantity of row is p, the described second average pond unit is used to be averaging every column element in each second matrix P obtained second sequence signatures indicate to be averaging, and p is positive integer, p>1.

14. according to any device of claim 11 to 13, which is characterized in that the first circulation neural network is in training When include that the first LSTM units, the first average pond unit and first train grader；

Described device further includes：

Second acquisition module, for obtaining the sample sequence data, the sample sequence data include：The first of sample of users The label data of sample behavior sequence, at least one the second sample behavior sequence and the sample of users for being associated with sample of users, The label data is the label after being labeled to the credit of the sample of users, and the association sample of users is in social network There are associated other users with the sample of users for network；

5th computing module, the first sample behavior sequence input described first for obtaining second acquisition module LSTM units obtain training characteristics sequence；

6th computing module, first is averaged described in the training characteristics sequence inputting for obtaining the 5th computing module Pond unit obtains the expression of training sample sequence signature；

7th computing module, the training sample sequence signature for obtaining the 6th computing module indicate described in input First training grader, obtains the first prediction data；

Judgment module, first prediction data for obtaining the 7th computing module and the label data substitute into the One loss function, judges whether the first-loss function converges to minimum；

Module is adjusted, for when the judgment module judges that the first-loss function does not converge to minimum, utilizing Error backpropagation algorithm is adjusted the parameter of the first LSTM units, until the first-loss function convergence to pole Small value；

Determining module, for when the judgment module judges the first-loss function convergence to minimum, after adjustment The parameters of the first LSTM units be determined as the model parameter of the first circulation neural network.

15. according to the devices described in claim 11, which is characterized in that the stack self-encoding encoder includes k layers and is limited Bohr hereby Graceful machine RBM, each RBM includes input layer and hidden layer, and k is positive integer；

The third computing module, including：

Concatenation unit, the First ray character representation and described second for obtaining first computing module calculate mould Second sequence signature expression that block obtains is spliced, and the sequence signature for obtaining the user indicates；

Input unit, the sequence signature of the user for obtaining the concatenation unit indicate to input the stack own coding Device, the input of the input layer of the i+1 layer RBM of the stack self-encoding encoder are the defeated of the hidden layer of i-th layer of RBM Go out, i is positive integer, i<i+1<k；

First determination unit derives table for the assemblage characteristic by the output of the hidden layer of the kth layer RBM as the user Show.

16. the device according to claim 11 or 15, which is characterized in that the stack self-encoding encoder includes k in training The layer training graders of limited Boltzmann machine RBM and second, k is positive integer；

Described device further includes：

Input module, the sequence signature for mixing the sample with family indicate the input stack self-encoding encoder, the sample of users Sequence signature expression is calculated according to the sample sequence data；

First training module, in the pre-training stage, carrying out unsupervised learning training to RBM described in every layer respectively, obtaining pre- RBM parameters after training；

Second training module, in the accurate adjustment stage, supervision have been carried out in conjunction with the second training grader RBM described to k layers Learning training, the supervised learning training is for the RBM ginsengs after adjusting the pre-training that first training module obtains Number.

17. device according to claim 16, which is characterized in that each RBM includes input layer in training, hides Layer and output layer；

First training module, including：

5th computing unit, for when being trained to i-th layer of RBM, predetermined characteristic to be indicated described in i-th layer of input The input layer of RBM, according to the output of i-th layer of RBM of the i-th weight matrix and the calculating of the i-th bias vector of i-th layer of RBM The data of layer；

First judging unit, the output for obtaining the data of the input layer of i-th layer of RBM and the 5th computing unit The data of layer substitute into the second loss function, judge whether second loss function converges to minimum；

The first adjustment unit judges that second loss function does not converge to minimum for working as first judging unit When, i-th weight matrix and i-th bias vector are adjusted, after the data of the input layer of i-th layer of RBM and adjustment The data of output layer of i-th layer of RBM substitute into second loss function, until second loss function converges to pole Small value；

Second determination unit, for when first judging unit judges that second loss function converges to minimum, By after adjustment i-th weight matrix and i-th bias vector be determined as the RBM parameters after the pre-training；

Wherein, input of the output of the hidden layer of i-th layer of RBM as the input layer of the i+1 layer RBM, i are just whole Number, the predetermined characteristic indicates it is that the sequence signature of the sample of users indicates that i is more than predetermined characteristic table when 1 when i is 1 Show the output for the hidden layer for being (i-1)-th layer of RBM, i<i+1<k.

18. device according to claim 16, which is characterized in that second training module, including：

6th computing unit indicates input for deriving the output of the hidden layer of the kth layer RBM as training assemblage characteristic To the second training grader, the second prediction data is obtained；

Second judgment unit judges institute for second prediction data and the label data to be substituted into third loss function State whether third loss function converges to minimum；

Second adjustment unit judges that the third loss function does not converge to minimum for working as the second judgment unit When, the weight matrix and bias vector of the hidden layer of RBM described in every layer are adjusted using error backpropagation algorithm, until The third loss function converges to minimum；

Third determination unit, for when the second judgment unit judges that the third loss function converges to minimum, The weight matrix of the hidden layer of every layer of RBM after adjustment and bias vector are determined as to the model of the stack self-encoding encoder Parameter.

19. device according to claim 16, which is characterized in that the first circulation neural network includes the first LSTM mono- Member and the first average pond unit；The second circulation neural network includes the 2nd LSTM units and the second average pond unit； The sample sequence data include the sample of users first sample behavior sequence, it is at least one association sample of users second The label data of sample behavior sequence and the sample of users, the label data are to the credit of the sample of users into rower Label after note, the association sample of users are that there are associated other users with the sample of users in social networks；

Described device further includes：

8th computing module obtains the sample for the first sample behavior sequence to be inputted the first LSTM units The fisrt feature sequence of user；

9th computing module, first is averaged described in the fisrt feature sequence inputting for obtaining the 8th computing module Pond unit obtains the expression of first sample sequence signature；

Tenth computing module obtains the association for the second sample behavior sequence to be inputted the 2nd LSTM units The parameter of the second feature sequence of sample of users, the 2nd LSTM units is identical as the parameter of the first LSTM units；

11st computing module, second is flat described in the second feature sequence inputting for obtaining the tenth computing module Equal pond unit, obtains the second sample sequence character representation；

Concatenation module, the first sample sequence signature for obtaining the 9th computing module indicate and the described 11st The second sample sequence character representation that computing module obtains is spliced, and the sequence signature table of the sample of users is obtained Show, the sequence signature expression of the sample of users is the input of the stack self-encoding encoder；

Wherein, when the quantity of the second sample behavior sequence is one, the described second average pond unit is used for described Second feature sequence is averaging；When the quantity of the second sample behavior sequence is p, the described second average pond unit is used It is averaging in the p being averaging to each second feature sequence the second sample sequence character representations, p is just whole Number, p>1.

20. according to any device of claim 11 to 19, which is characterized in that first acquisition module, including：

First map unit, for obtaining n vector included by the first behavior sequence of the user, by described n vector It is mapped in same Value space, n is positive integer；

Second map unit, for obtaining m vector included by the second behavior sequence of the association user, by the m In DUAL PROBLEMS OF VECTOR MAPPING to the same Value space, m is positive integer.