CN108876586A - A kind of reference point determines method, apparatus and server - Google Patents
A kind of reference point determines method, apparatus and server Download PDFInfo
- Publication number
- CN108876586A CN108876586A CN201710330665.2A CN201710330665A CN108876586A CN 108876586 A CN108876586 A CN 108876586A CN 201710330665 A CN201710330665 A CN 201710330665A CN 108876586 A CN108876586 A CN 108876586A
- Authority
- CN
- China
- Prior art keywords
- head portrait
- network
- feature
- user
- network head
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
- G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/03—Credit; Loans; Processing thereof
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- General Business, Economics & Management (AREA)
- Finance (AREA)
- Data Mining & Analysis (AREA)
- Accounting & Taxation (AREA)
- Educational Administration (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Technology Law (AREA)
- Game Theory and Decision Science (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present invention provides that a kind of reference point determines method, apparatus and server, this method include:Obtain the network head portrait of target user;Determine the characteristics of image of the network head portrait;The head portrait attribute of the network head portrait is determined according to described image feature;According to the reference sub-model of pre-training, reference point corresponding with the behavioural characteristic of the head portrait attribute and the target user is determined;Wherein, the reference sub-model is obtained according at least to the head portrait attribute training of the behavioural characteristic and network head portrait of positive sample user and negative sample user, and the creditworthiness of positive sample user is higher than negative sample user.The embodiment of the present invention is determining reference timesharing, has taken the behavioural characteristic of user and the head portrait attribute of network head portrait into consideration, so that the determination factor of reference point is more fully, helps to promote the determining effect of final reference point.
Description
Technical field
The present invention relates to technical field of data processing, specifically design a kind of reference point and determine method, apparatus and server.
Background technique
Reference point is that the score value of user credit degree embodies, and can indicate the Default Probability of user;Reference has been divided not at present
It is used alone in credit field, also has in other such as shared economy, user's evaluation, information recommendation field and be widely applied.
With the quickening of reference marketization paces, also constantly extended using the field of reference point;Therefore how to optimize
The method of determination of reference point is of great significance, this is also always the technical point of those skilled in the art's research.
Summary of the invention
In view of this, the embodiment of the present invention, which provides a kind of reference point, determines method, apparatus and server, to optimize reference point
Method of determination.
To achieve the above object, the embodiment of the present invention provides the following technical solutions:
A kind of determining method of reference point, including:
Obtain the network head portrait of target user;
Determine the characteristics of image of the network head portrait;
The head portrait attribute of the network head portrait is determined according to described image feature;
According to the reference sub-model of pre-training, determination is corresponding to the behavioural characteristic of the head portrait attribute and the target user
Reference point;Wherein, behavioural characteristic and network head of the reference sub-model according at least to positive sample user and negative sample user
The head portrait attribute training of picture obtains, and the creditworthiness of positive sample user is higher than negative sample user.
The embodiment of the present invention also provides a kind of reference point determining device, including:
Head portrait obtains module, for obtaining the network head portrait of target user;
Characteristics of image determining module, for determining the characteristics of image of the network head portrait;
Head portrait attribute determination module, for determining the head portrait attribute of the network head portrait according to described image feature;
Reference divides determining module, for the reference sub-model according to pre-training, determining and the head portrait attribute and the mesh
Mark the corresponding reference of behavioural characteristic point of user;Wherein, the reference sub-model is used according at least to positive sample user and negative sample
The behavioural characteristic at family and the head portrait attribute training of network head portrait obtain, and the creditworthiness of positive sample user is higher than negative sample user.
The embodiment of the present invention also provides a kind of server, including:At least one processor and at least one processor;It is described
Memory is stored with program, and the processor calls the program of the memory storage, and described program is used for:
Obtain the network head portrait of target user;
Determine the characteristics of image of the network head portrait;
The head portrait attribute of the network head portrait is determined according to described image feature;
According to the reference sub-model of pre-training, determination is corresponding to the behavioural characteristic of the head portrait attribute and the target user
Reference point;Wherein, behavioural characteristic and network head of the reference sub-model according at least to positive sample user and negative sample user
The head portrait attribute training of picture obtains, and the creditworthiness of positive sample user is higher than negative sample user.
Based on the above-mentioned technical proposal, server can obtain the network head portrait of target user, determine the figure of the network head portrait
As feature, and determine according to described image feature the head portrait attribute of the network head portrait;To pass through the combination user of pre-training
Behavioural characteristic and head portrait attribute reference sub-model, can be when the reference point for carrying out target user determines, by target user's
The head portrait attribute of behavioural characteristic and network head portrait is imported in the reference sub-model, to be handled by reference sub-model, is determined
The reference of target user point, identified reference point have taken the behavioural characteristic of target user and the head portrait category of network head portrait into consideration
Property, so that the determination factor of reference point is more fully, help to promote the determining effect of final reference point.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is the connection schematic diagram of server and business platform;
Fig. 2 is the hardware block diagram of server provided in an embodiment of the present invention;
Fig. 3 is the flow chart that reference provided in an embodiment of the present invention point determines method;
Fig. 4 is the schematic diagram of network head portrait;
Fig. 5 is the schematic diagram of head portrait subject classification;
Fig. 6 is the method flow diagram for determining the CNN feature of network head portrait;
Fig. 7 is that the determining schematic diagram of reference point is realized by CNN;
Fig. 8 is the schematic diagram of the CNN network of AlexNet model structure;
Fig. 9 is that training provided in an embodiment of the present invention obtains the method flow diagram of head portrait subject classification model;
Figure 10 is the training schematic diagram of head portrait subject classification model;
Figure 11 is that training provided in an embodiment of the present invention obtains the method flow diagram of reference sub-model;
Figure 12 is the structural block diagram that reference provided in an embodiment of the present invention divides determining device;
Figure 13 is another structural block diagram that reference provided in an embodiment of the present invention divides determining device;
Figure 14 is another structural block diagram that reference provided in an embodiment of the present invention divides determining device.
Specific embodiment
It was found by the inventors of the present invention that the behavior that the reference of user point is mainly based upon user determines, it can such as pass through acquisition
The behavioural characteristic of user predicts the Default Probability of user, determines the reference point of user;Although the behavior by user can
The reference point of user is evaluated, but the simple behavior based on user carries out the determination of reference point, the determination factor of reference point is then
Seem more single;Therefore the present inventor considers how that other factors that can influence user's reference point are added, to mention
The comprehensive of the determination factor of reference point is risen, realizes the optimization of the method for determination to reference point.
It was found by the inventors of the present invention that the network head portrait of user (in the application of social attribute, (such as answer such as user by instant messaging
With) in be arranged network head portrait, the head portrait etc. that for another example user uploads in websites such as forums) be that user identifies oneself in a network
Image is the first impression shown in a network to the external world, contains various dimensions very rich in the network head portrait of user
Information, and generally there is incidence relation to a certain degree with user's personality, animation in the network head portrait of user, and the property of user
Lattice, animation again to a certain extent can have an impact the reference of user;Therefore by the network head portrait of user, to being based on
User behavior determines that the mode of reference point is supplemented, and will be of great significance, and can make the determination factor of reference point more
Be it is comprehensive, help to be promoted final reference point and determine effect.
Based on this thinking, the present inventor proposes to combine the network head portrait of user, determines to the behavior based on user
The mode of reference point is supplemented, the sign for being optimized with the method for determination to reference point, and being provided through the embodiment of the present invention
Letter point determines that method is implemented.
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
Reference provided in an embodiment of the present invention point determines that method can be applied in server, which can be specially
What is be arranged divides determining service equipment for carrying out reference;The server is also possible to the service equipment in server cluster, should
Server cluster may be considered the collection for realizing the server of a certain main business (such as social activity, e-commerce, Third-party payment)
Group, can be by carrying out the extension of service equipment, alternatively, implementing in original server upper set present invention in the server cluster
The reference point that example provides determines function;
As shown in Figure 1, the server can collect the data of at least one business platform, (Fig. 1 show more than two
Business platform, but the practical business platform that also can be set is one or more, specific business platform quantity can be according to practical feelings
Condition setting), the behavioral data of user is collected into from least one business platform, which can be corresponding with
At least one type of service, a type of service can be supported by one or more business platforms;
Optionally, in the embodiment of the present invention, the type of service of business platform can be such as social category business platform, video class industry
Be engaged in platform, using downloading business platform, e-commerce business platform etc., specifically may be set according to actual conditions;
Optionally, which can share an account system, i.e. user can be used same account not
Same business platform, which is realized, to be logged in, if at least one business platform can share the account system of social platform, so that user can
It is realized and is logged in different business platforms using social account, then the embodiment of the present invention can be identified same by the social account of user
One user and is collected in the behavioral data of different business platform;Further, same user at least one business platform
Network head portrait can be identical, the network head portrait such as uploaded in social platform using user;
Certainly, which may not also share an account system, as at least one business is flat
There are business platforms not to connect in the business platform of banking institution, government department involved in platform or at least one business platform
Enter shared account system standard (as there is the e-commerce platform for not accessing shared social account system), then the present invention is real
Applying example can be used the identity such as identification card number, the cell-phone number of user, identify same user in the behavior number of different business platform
According to;Further, it may be determined that the main business platform at least one business platform, the network uploaded with user in main business platform
Head portrait conduct, the embodiment of the present invention carry out the determining Replacement Factor of reference point.
The embodiment of the present invention can load corresponding program in the server, realize that reference provided in an embodiment of the present invention point determines
Method, the program can be stored by the memory in server, and called and implemented by processor.
Optionally, Fig. 2 shows the hardware block diagrams of server provided in an embodiment of the present invention, referring to Fig. 2, the service
Device may include:At least one processor 1, at least one communication interface 2, at least one processor 3 and at least one communication are total
Line 4;
In embodiments of the present invention, processor 1, communication interface 2, memory 3, communication bus 4 quantity be at least one,
And processor 1, communication interface 2, memory 3 complete mutual communication by communication bus 4;Obviously, processor shown in Fig. 2
1, the communication connection signal of communication interface 2, memory 3 and communication bus 4 is only optional;
Optionally, communication interface 2 can be the interface of communication module, such as the interface of gsm module;
Processor 1 may be a central processor CPU or specific integrated circuit ASIC (Application
Specific Integrated Circuit), or be arranged to implement the integrated electricity of one or more of the embodiment of the present invention
Road.
Memory 3 may include high speed RAM memory, it is also possible to further include nonvolatile memory (non-volatile
Memory), a for example, at least magnetic disk storage.
Wherein, memory 3 is stored with program, and the program that processor 1 calls memory 3 to be stored realizes the embodiment of the present invention
The reference of offer point determines method.
Below from the angle of server, method, which is introduced, to be determined to reference provided in an embodiment of the present invention point, hereafter
The reference of description point determines method content, and in combination with the network head portrait of user, the side of reference point is determined to the behavior based on user
Formula is supplemented, in the specific implementation, can be real by the corresponding program of server load to optimize to reference point method of determination
It is existing.
Fig. 3 is the flow chart that reference provided in an embodiment of the present invention point determines method, and this method can be applied to server, is joined
According to Fig. 3, this method may include:
Step S100, the network head portrait of target user is obtained.
Target user can be the determining user of the pending reference of the embodiment of the present invention point, and the embodiment of the present invention can will be any
User is as target user, so that the reference provided through the embodiment of the present invention point determines that method carries out the determination of reference point, it is real
Now the reference of each user point determines.
Optionally, the network head portrait of target user can be target user and exist in uploads such as website, social application platforms
The image of tag image on network can represent target user, and the network head portrait of target user can be with target user in net
Network (such as website, social application platform) setting is associated with account name, alias;Fig. 4 shows a kind of signal of network head portrait,
It can refer to;
Optionally, network head portrait can be the image of two dimensional form, be also possible to the image of three dimensional form, and specific view is practical
Depending on situation.
Step S110, the characteristics of image of the network head portrait is determined.
Optionally, the embodiment of the present invention can pass through CNN (Convolutional Neural Network, convolutional Neural net
Network) the CNN feature that determines the network head portrait, it is used using identified CNN feature as characteristics of image;Obviously, it removes and uses CNN
Outside, the embodiment of the present invention can also use other image characteristics extraction models, and image spy is extracted from the network head portrait
Sign, correspondingly, the form of characteristics of image is not limited to CNN feature, but can be according to used image characteristics extraction model
Corresponding adjustment.
Step S120, the head portrait attribute of the network head portrait is determined according to described image feature.
Optionally, head portrait attribute may be considered head portrait subject description corresponding to head portrait, i.e., the head of the described network head portrait
As attribute may be considered, head portrait subject description corresponding to the network head portrait.
Optionally, head portrait subject description corresponding to head portrait, can be corresponding in each head portrait subject classification by head portrait
Probability realizes that head portrait is known as in the corresponding probability of each head portrait subject classification, and head portrait is in each head portrait subject classification
Probability distribution;I.e. optional, head portrait subject description corresponding to the network head portrait can be, and the network head portrait is in each head
As the corresponding probability of subject classification, i.e., probability distribution of the described network head portrait in each head portrait subject classification.
Optionally, determine that the network head portrait in the probability distribution of each head portrait subject classification, can pass through pre-training
Head portrait subject classification model realization;Head portrait subject classification model can by machine learning algorithm (such as Softmax sorting algorithm),
The characteristics of image of the image of each head portrait subject classification of training is realized, as the embodiment of the present invention collects each head portrait subject classification
Image, image collected by a head portrait subject classification can be one or more, to analyze collected each head portrait
The characteristics of image of the image of subject classification, using machine learning algorithms such as Softmax sorting algorithms, to each head portrait subject classification
The characteristics of image of image be trained, determine the head portraits subject classification model such as Softmax disaggregated model, the Softmax points
Class model can predict the image in the probability of each head portrait subject classification based on the characteristics of image of an image;
It is worth noting that, a head portrait the corresponding probability of each head portrait subject classification and be 1, such as the present invention implement
The settable m head portrait subject classification of example, head portrait can be expressed as (p in the corresponding probability of each head portrait subject classification1,p2,p3,
p4,p5,p6,…,pi,…,pm), wherein piIt is considered the probability of head portrait head portrait subject classification at i-th, 0 < i <=m,
Then
Optionally, the number quantity and form of set head portrait subject classification can be arranged according to the actual situation;Such as it can be with
According to statistics by the subject classification of head portrait be divided into plant, people, microorganism, animal and sports, artificial facility, natural objects and
Several level-one classifications such as other, each level-one class can be subdivided into multiple second level classifications again now, as shown in Figure 5;Obviously, Fig. 5
The set-up mode of the subject classification of shown head portrait is only optionally that the class of multiple peers can also be directly arranged in the embodiment of the present invention
Subject classification of the mesh as head portrait;For the setting of specific head portrait subject classification, can adjust according to the actual situation.
Step S130, according to the reference sub-model of pre-training, the determining row with the head portrait attribute and the target user
It is characterized corresponding reference point.
Optionally, reference sub-model can be the mould of the reference for assessing user point of training in advance of the embodiment of the present invention
Type, different with the training process of traditional reference sub-model to be, the embodiment of the present invention can be according at least to the behavior spy of user
It seeks peace head portrait attribute, carries out the training of reference sub-model, so that the head portrait attribute of user is fused in reference sub-model, so that
It is contemplated that more comprehensively influencing the factor of user's reference in the training process of reference sub-model, reference sub-model is promoted
Overall effect;Correspondingly, the reference sub-model trained with this can combine user in the reference point assessment for carrying out user
Head portrait attribute realize that the reference point of user determines;
Optionally, when carrying out the training of reference sub-model, the embodiment of the present invention can be according to the historical behavior of user, from more
Determine that behavior promise breaking is less in a user, the higher user of creditworthiness determines that behavior is broken a contract as positive sample user
More, the lower user of creditworthiness is as negative sample user;Optionally, the determination of positive sample user and negative sample user can be with
It is realized by the historical behavior of manual analysis user, certainly, if there is corresponding automatic algorithms can be realized, the embodiment of the present invention
Also it is not precluded in a manner of automatic algorithms, positive sample user and negative sample user is determined from multiple users;
In turn, the embodiment of the present invention can determine the behavioural characteristic of positive sample user and the head portrait attribute of network head portrait, and
The behavioural characteristic of negative sample user and the head portrait attribute of network head portrait are merged to be trained with machine learning method
There is the reference sub-model of head portrait attribute;Optionally, the behavioural characteristic of user can be analyzed to obtain by the behavioral data of user.
Optionally, the embodiment of the present invention can not be related to the improvement to machine learning method, but utilize machine learning side
When method, by adjusting the feature that model training uses, so that the feature that training uses also is melted in addition to the behavioural characteristic comprising user
Share the head portrait attribute at family, thus for it is subsequent obtain fusion there is the reference sub-model of head portrait attribute to provide possibility.
As can be seen that by the behavioural characteristic of the combination user of pre-training and the reference sub-model of head portrait attribute, the present invention
Embodiment, can be by the head portrait category of the behavioural characteristic of target user and network head portrait when the reference point for carrying out target user determines
Property, it imports in the reference sub-model, to handle by reference sub-model, determines the reference point of target user, it is identified
Reference point has taken the behavioural characteristic of target user and the head portrait attribute of network head portrait into consideration, so that the determination factor of reference point is more
Be it is comprehensive, help to be promoted final reference point and determine effect.
Optionally, the reference of identified target user point, which can be applied, is carrying out credit evaluation (generally to target user
, reference point is higher, and credit evaluation amount is higher), (reference of such as target user point is higher, then target user institute for user's evaluation
The confidence level of the evaluation of progress is higher), economy is shared, (reference of such as target user point is higher, then target user sends out for information recommendation
The confidence level of the information of cloth is higher, can preferentially be recommended) etc. fields.
Optionally, the embodiment of the present invention can determine the network head when determining the characteristics of image of network head portrait by CNN
The CNN feature of picture;It should be noted that CNN is one kind of neural network, there is outstanding table on processing image problem
Existing, its artificial neuron can respond the image pixel elements in certain coverage area, and CNN model is most commonly used at image
The model of reason;Compared to traditional artificial nerve network model, CNN possesses more hidden layers, distinctive convolution sum pond
Operation to image handled higher efficiency, image can be considered in digital image processing field pixel to
Amount, for example the image of a 100*100 size can be converted into one 10000 vector;
If directly realizing the extraction of characteristics of image with traditional artificial nerve network model, it is assumed that the number of hidden layer is also
10000, then their weight connected will be 108A, more parameters makes the various parameters of neural network very in this way
Difficulty is trained to, while the relevant information on image space between surrounding neighbors can be lost if image is pulled into vector;
And CNN is to share two kinds of reasonable forms by local receptor field and weight to realize, has pole in the quantity for reducing parameter
Big advantage, therefore it is preferable to use CNN for the embodiment of the present invention, realize the extraction of characteristics of image.
The embodiment of the present invention can pass through multiple convolutional layers of CNN when determining the CNN feature of network head portrait by CNN
Process of convolution is carried out to network head portrait, and tieing up as a result, forming setting for multiple convolutional layer process of convolution is connected by full articulamentum
The feature vector of degree obtains the CNN feature of network head portrait;Wherein, the object of the first convolutional layer process of convolution is network head portrait, non-
The object of first convolutional layer process of convolution is the result after upper one layer of convolutional layer process of convolution;
Optionally, the method flow for determining the CNN feature of network head portrait as shown in Figure 6, it is special in the CNN for carrying out network head portrait
Really timing, the process may include sign:
Step S200, by network head portrait import CNN first layer convolutional layer, by first layer convolutional layer to network head portrait into
Row process of convolution.
Step S210, the convolution processing result of first layer convolutional layer is imported into next layer of convolutional layer and carries out process of convolution, and
So that in next layer of convolutional layer process of convolution one layer of convolutional layer convolution processing result, until all convolutional layers carry out pulleying
Product processing.
Optionally, the convolution processing result of a convolutional layer may be considered the Feature Map, a Feature of multiple groups
Map indicates one group of various dimensions feature after the convolutional layer process of convolution;
Optionally, process of convolution of the embodiment of the present invention with layer-by-layer convolutional layer, the process of convolution knot of next layer of convolutional layer
The corresponding characteristic dimension of fruit can be lower than the corresponding characteristic dimension of convolution processing result of upper one layer of convolutional layer, i.e., with layer-by-layer volume
The process of convolution of lamination, the corresponding characteristic dimension of convolution processing result will realize dimensionality reduction;And the process of convolution of each layer of convolutional layer
As a result the group number of corresponding dimensional characteristics can be set according to the actual conditions of each convolutional layer;For example, the volume of second layer convolutional layer
Product processing result is the feature of 256 groups of x*y dimension, and the convolution processing result of third layer convolutional layer is 384 groups of x1*y1The feature of dimension, then
x1< x, y1< y realizes the dimensionality reduction of the corresponding characteristic dimension of convolution processing result, and 256 groups referred here to, 384 groups of group number are only
It is citing parameter, specific value can adjust setting according to the actual situation, and not have restriction effect.
Step S220, the convolution processing result of final layer convolutional layer is imported into full articulamentum, connection forms setting dimension
Feature vector obtains the CNN feature of network head portrait.
Optionally, the convolution processing result of final layer convolutional layer is also the various dimensions feature of multiple groups, is conducted into full connection
After layer, the various dimensions feature of the multiple groups can be connected to become the feature vector of setting dimension (such as one-dimensional) by full articulamentum, obtain net
The CNN feature of headstall picture.
Optionally, the embodiment of the present invention can also be arranged Softmax layers in CNN, and the Softmax layers can have the present invention
The function for the Softmax disaggregated model that embodiment provides, as CNN is arranged in the Softmax disaggregated model that will be trained
Softmax layers, so as to the CNN feature for the network head portrait for exporting full articulamentum, Softmax layers is imported into, obtains network head
As the probability distribution in each head portrait subject classification;
As shown in fig. 7, the network head portrait of target user can be handled by the convolutional layer of CNN network and full articulamentum is handled,
The CNN feature of network head portrait is obtained, which imports Softmax layers, and network head portrait can be determined in each head portrait theme
The probability distribution of classification;Thus probability distribution and mesh of the network head portrait of combining target user in each head portrait subject classification
The behavioural characteristic for marking user by reference sub-model, can determine the reference point of target user.
It should be noted that probability distribution of the network head portrait in each head portrait subject classification, it can be with network head portrait institute body
Existing user's personality, there are incidence relations for animation, therefore can indirectly influence the reference of user;To be obtained in big data analysis
The statistical law explanation arrived, if user uses oneself or household's photo as network head portrait, which tends to certainly
Centered on me, there is stronger self-consciousness;If user uses cartoon picture as network head portrait, which often compares
Miss old times or old friends, is easy by extraneous things or so, possible more satisfactoryization of thought, but the possible creative consciousness of such user is very strong;If
User uses the scenery is beautiful to scheme as network head portrait, which is often people's maturation, there is the ability of stronger processing problem;If user
Use child or baby's picture etc. as network head portrait, which is likely to married user, and happiness of living is happy;Compare again
Such as, if containing contact methods, the users such as telephone number in the network head portrait of user may be to be engaged in using the network platform
Transaction.
As it can be seen that not only shown the personality of individual subscriber in the network head portrait of user, also contained subscriber household, love and marriage,
The animations information such as occupation, emotion;It is the displaying of individual subscriber image, is that user shows on the networks such as social platform
Extraneous first impression, contains the information of various dimensions very rich, by these Information applications into user's reference, it will help
In the building effect for promoting reference sub-model, so that determining the determination factor of reference sub-model more fully.
Optionally, network head portrait is only the one of the head portrait attribute of network head portrait in the probability distribution of each head portrait subject classification
Kind preferred form;As long as user's personality that the head portrait attribute of network head portrait can be embodied with network head portrait, animation exist
Incidence relation, the head portrait attribute of other forms is not precluded in the embodiment of the present invention, such as in the base based on color reaction user's personality
On plinth, it can also be distributed using the color character of network head portrait as head portrait attribute.
Optionally, when carrying out the building of CNN, the models such as AlexNet, VGG, ResNet are can be used in the embodiment of the present invention
Structure is realized;
Here, AlexNet is that Alex Krizhevsky etc. is based on GPU computing platform proposition AlexNet model, million
On the ImageNet data acquisition system of magnitude, effect is significantly more than traditional method, is promoted from traditional 70% to more than 80% more.
AlexNet training is based on ImageNet image data set, and big data and GPU highly-parallel calculate so that deep learning is led in vision
The application in domain is possibly realized;AlexNet is by deepening network and introducing the innovatory algorithms such as ReLU, Dropout.AlexNet is deep
Degree study is represented in the milestone formula of visual development.
It is go deeper that VGG and GoogLeNet these two types model structure, which has a common feature,.GoogLeNet is by Top5
Lower error rate be one 22 layers of depth convolutional neural networks to 6.67%, GoogLeNet;Increase for convenience and repairs
Change, GoogLeNet uses modular structure;Network uses Average Pooling finally to replace full articulamentum.For
Facilitate finetune, can also finally add a full articulamentum;In order to avoid over-fitting, still used in network
Dropout;In order to avoid gradient disappearance, network adds additional the SoftMax of 2 auxiliary for conducting gradient forward.
VGG inherits some frames of Lenet and Alexnet, it is more similar with Alexnet frame, and VGG is also 5
The convolution of a Group, 2 layers of full connection characteristics of image, 1 layer of full link sort feature can regard as Alexnet in total 8 as
A part.According to the difference configuration in preceding 5 convolution Group, each Group, VGG gives this five kinds configurations of A~E, convolution
The number of plies is incremented by from 8 to 16, with the intensification step by step of convolutional layer from 8 to 16, is also had arrived at accurately by deepening the convolution number of plies
The bottleneck that rate is promoted.
Level is referred to 152 layers by ResNet, and Top-5 accuracy rate is reduced to 3.57%.ResNet is based on depth residual error
The irregular network application of depth to field of image recognition is significantly reduced the difficulty of trained deeper time neural network by learning framework
Degree, also makes accuracy rate be significantly improved.The main innovation of ResNet in residual error network, substantially solution level it is deep when
Wait the problem that can not be trained.The network that it has used for reference Highway Network thought is equivalent to the special channel of opening in side and makes
The output that can go directly is inputted, and the target optimized is become the difference of output and input by original fitting output.
The extraction of CNN feature for ease of understanding, the embodiment of the present invention by taking the CNN network of the model structure of AlexNet as an example,
Illustrate the extraction process of CNN feature;As shown in figure 8, the CNN network of the model structure of AlexNet may include 5 layers of convolutional layer
(the full articulamentum of Conv1 to Conv5) and full articulamentum, certain CNN network is also Softmax layers settable later;Optionally,
Softmax disaggregated model can also be placed in except CNN network by the embodiment of the present invention;
Based on shown in Fig. 8, the example of CNN feature extraction can following process, it is notable that it is following be related to it is specific
Numerical value is merely illustrative of, and can according to circumstances be adjusted in actual use, but generally as the convolution of layer-by-layer convolutional layer
Processing, the corresponding characteristic dimension of convolution processing result will realize dimensionality reduction;
Step 1:The picture of network head portrait is divided into m n*n fritter, the first convolutional layer Conv1 by Conv1 stage, Conv1
Convolution kernel using sliding window mechanism realize, convolution algorithm is carried out on the picture of network head portrait, specifically can be by the first convolution
96 neurons of layer carry out process of convolution to network head portrait respectively, each neuron handles to obtain a s*t dimensional feature (such as
Assuming that the frame size of sliding window is k*k, the picture size of network image is l*z, then s=l-k+1, t=z-k+1), one
S*t dimensional feature is placed in a figure, has reformed into the Feature Map (FM) of a s*t, so that one is obtained 96 FM, one
Open the corresponding one group of s*t dimensional characteristics of FM;Down-sampling operation (Down Sampling) is carried out later, by 96 groups in Feature Map
S*t dimensional characteristics are transformed to 96 groups of p*q, and (p*q is that the dimension after sampling, specific rules and dimension numerical value are true according to sampling rule
It is fixed) feature, wherein p < s, q < t;Feature after down-sampling is input to normalization layer, then normalizes result input
Conv2;
Step 2:Conv2 stage, Conv2 carry out convolution algorithms to 96 groups of p*q dimensional features respectively, generates 256 groups (256
Numerical value is parameter, can tune up or turn down, can be determined by test) i*j dimensional feature, it puts them in a figure,
Reform into the Feature Map (FM) of an i*j, 256 altogether;Down-sampling operation (Down Sampling) is carried out later,
By 256 groups of i*j dimensional characteristics in Feature Map be transformed to 256 groups of x*y (x*y be sampling after dimension, specific rules and
Dimension numerical value is determined according to sampling rule) feature, wherein x < i, y < j;Feature after down-sampling is input to normalization layer,
Then normalization result inputs Conv3;
Step 3:Conv3 stage, Conv3 carry out convolution algorithms to 256 groups of x*y dimensional features respectively, generates 384 groups (384
Numerical value is parameter, can tune up or turn down, can be determined by test) x1*y1Dimensional feature is put them in a figure,
An x is reformed into1*y1Feature Map (FM), 384 altogether, then input Conv4;
Step 4:In the Conv4 stage, Conv4 is respectively to 384 groups of x1*y1Dimensional feature progress convolution algorithm, generation 384 (384
Numerical value is parameter, can tune up or turn down, can be determined by test) group x2*y2Dimensional feature puts them on a figure
In, reform into an x2*y2Feature Map (FM), 384 altogether, then input Conv5;
Step 5:In the Conv5 stage, Conv5 is respectively to 384 groups of x2*y2Dimensional feature progress convolution algorithm, generation 256 (256
Numerical value is parameter, can tune up or turn down, can be determined by test) group x3*y3Dimensional feature puts them on a figure
In, reform into an x3*y3Feature Map (FM), 256 altogether, then input full articulamentum;
Step 6:Full access phase, full articulamentum are attached the output result of Conv5, form the feature of setting dimension
Vector obtains the CNN feature of network head portrait.
It is handled further, it is possible to which the CNN feature that Softmax layers can export full articulamentum is arranged, determines network head
As the probability distribution in each head portrait subject classification.
Optionally, the embodiment of the present invention can determine the characteristics of image of network head portrait by training head portrait subject classification model
Corresponding, network head portrait realizes the determination of the head portrait attribute of network head portrait in the probability of each head portrait subject classification;
Optionally, Fig. 9 shows training provided in an embodiment of the present invention and obtains the method flow of head portrait subject classification model
Figure, referring to Fig. 9, this method may include:
Step S300, the image of each head portrait subject classification is obtained.
Optionally, a head portrait subject classification can correspond to the one or more images of acquisition.
Step S310, the corresponding characteristics of image of image of each head portrait subject classification is obtained.
Optionally, the embodiment of the present invention by method shown in Fig. 6, can get each head portrait subject classification image it is corresponding
CNN feature, for each image of each head portrait subject classification, the corresponding acquisition for carrying out CNN feature.
Step S320, according to machine learning algorithm, the characteristics of image of the image of each head portrait subject classification of training is obtained to the end
As subject classification model.
Optionally, as shown in Figure 10, the settable multiple head portrait subject classifications of the embodiment of the present invention, collect each head portrait theme
Classify corresponding image, and determine the CNN feature of the corresponding image of each head portrait subject classification, to pass through Softmax classification calculation
Method, training obtain Softmax disaggregated model, realize the acquisition of head portrait subject classification model.
Optionally, the embodiment of the present invention can be assessed by the reference sub-model for combining the head portrait attribute training of user to obtain
The reference of user point;Optionally, Figure 11 shows training provided in an embodiment of the present invention and obtains the method flow of reference sub-model
Figure, referring to Fig.1 1, this method may include:
Step S400, positive sample user and negative sample user are determined, wherein the creditworthiness of positive sample user is higher than negative sample
This user.
Optionally, the embodiment of the present invention can according to the historical behavior of user, determined from multiple users behavior promise breaking compared with
Few, the higher user of creditworthiness determines that behavior promise breaking is more as positive sample user, and the lower user of creditworthiness makees
Be negative sample of users.
Step S410, the behavioural characteristic of each positive sample user and the head portrait attribute of network head portrait are obtained, as positive sample
Feature;And the behavioural characteristic of each negative sample user and the head portrait attribute of network head portrait are obtained, as negative sample feature.
Optionally, for each positive sample user, the embodiment of the present invention can be obtained according to the behavioral data of positive sample user
The behavioural characteristic for getting positive sample user gets the image of the network head portrait of positive sample user according to head portrait subject classification model
Feature is corresponding, network head portrait each head portrait subject classification probability distribution (a kind of optional form of head portrait attribute), thus
In conjunction with a positive sample user behavioural characteristic and network head portrait in the probability distribution of each head portrait subject classification, obtain the positive sample
The sample characteristics of user can obtain the positive sample feature that training uses hence for each positive sample with this processing is made per family;
The acquisition process of negative sample feature is similar with the acquisition process of positive sample feature, is only that object is become by positive sample user
Be negative sample of users, and the two can be cross-referenced.
Step S420, it according to the machine learning algorithm training positive sample feature and negative sample feature, obtains reference and divides mould
Type.
Optionally, machine learning algorithm used in the embodiment of the present invention can be such as support vector machines (SVM), logistic regression
(Logitic Regression) etc..
Reference provided in an embodiment of the present invention point determines that one of method applies example, is shown in Fig.12;User A is
The registration user of social application, when carrying out reference point assessment to user A, server (determination for reference point) can be from society
The network head portrait for handing over the behavioral data for getting user A in the server of application and user A to upload;
To which server can determine the CNN feature of the network head portrait of user A by CNN network, classified using Softmax
Model handles the CNN feature, determines probability of the network head portrait in each head portrait subject classification of user A;Meanwhile it taking
Business device can analyze the behavioral data of user A, identify the behavioural characteristic of user A;
In turn, server by the network head portrait of user A the probability and user A of each head portrait subject classification behavior
Feature imports reference sub-model, and (behavioural characteristic of reference sub-model combination user and the head portrait attribute of network head portrait are trained
To), evaluate the reference point of user A.
The head portrait attribute of the network head portrait of user is fused in the establishment process of reference sub-model by the present invention, so that reference
The determination factor divided is more fully, it will help promotes the overall effect of constructed reference sub-model;And it is current general
The network platform all containing the head image data of user, in conjunction with user the reference point realized of head image data determine will be provided with it is very strong general
Adaptive.
Reference point determining device provided in an embodiment of the present invention is introduced below, reference described below point determines dress
It sets and may be considered server, the reference point that embodiment provides to realize the present invention determines the program module being arranged needed for method;
Reference described below divides the content of determining device, can determine that the content of method corresponds to each other reference with the reference of upper description point.
Figure 12 is the structural block diagram that reference provided in an embodiment of the present invention divides determining device, which can be applied to service
Device, referring to Fig.1 2, the apparatus may include:
Head portrait obtains module 100, for obtaining the network head portrait of target user;
Characteristics of image determining module 200, for determining the characteristics of image of the network head portrait;
Head portrait attribute determination module 300, for determining the head portrait attribute of the network head portrait according to described image feature;
Reference divides determining module 400, determining with the head portrait attribute and described for the reference sub-model according to pre-training
The corresponding reference of the behavioural characteristic of target user point;Wherein, the reference sub-model is according at least to positive sample user and negative sample
The behavioural characteristic of user and the head portrait attribute training of network head portrait obtain, and the creditworthiness of positive sample user is used higher than negative sample
Family.
Optionally, described image feature can be CNN feature, correspondingly, characteristics of image determining module 200, for determining
The characteristics of image of the network head portrait, specifically includes:
Process of convolution is carried out to network head portrait by multiple convolutional layers of CNN, and the multiple by the connection of full articulamentum
Convolutional layer process of convolution as a result, formed setting dimension feature vector, obtain the CNN feature of network head portrait.
Optionally, characteristics of image determining module 200 carries out convolution to network head portrait for multiple convolutional layers by CNN
Processing, specifically includes:
The first layer convolutional layer that network head portrait is imported to CNN carries out at convolution network head portrait by first layer convolutional layer
Reason;
The convolution processing result of first layer convolutional layer is imported into next layer of convolutional layer and carries out process of convolution, and makes next layer
The convolution processing result of one layer of convolutional layer in convolutional layer process of convolution, until all convolutional layers carried out process of convolution;
Correspondingly, characteristics of image determining module 200, for connecting the multiple convolutional layer process of convolution by full articulamentum
As a result, specifically including:
The convolution processing result of final layer convolutional layer is imported into full articulamentum, connection forms the feature vector of setting dimension,
Obtain the CNN feature of network head portrait.
Optionally, the corresponding characteristic dimension of the convolution processing result of next layer of convolutional layer, lower than the volume of upper one layer of convolutional layer
The corresponding characteristic dimension of product processing result;I.e. with the process of convolution of layer-by-layer convolutional layer, the corresponding feature dimensions of convolution processing result
Degree will realize dimensionality reduction.
Optionally, head portrait attribute determination module 300, for determining the head portrait of the network head portrait according to described image feature
Attribute specifically includes:
According to the head portrait subject classification model of pre-training, determination exists with described image feature correspondingly, the network head portrait
The probability distribution of each head portrait subject classification;The head portrait subject classification model is corresponding according at least to the image of each head portrait subject classification
Characteristics of image training obtain.
Optionally, the image color feature distribution that head portrait also can be used in head portrait attribute is realized.
Optionally, Figure 13 shows another structural block diagram of reference point determining device provided in an embodiment of the present invention, in conjunction with
Shown in Figure 12 and Figure 13, which can also include:
Subject classification model construction module 500, for obtaining the image of each head portrait subject classification;Obtain each head portrait theme
The corresponding characteristics of image of the image of classification;According to Softmax sorting algorithm, the image of the image of each head portrait subject classification of training
Feature obtains Softmax disaggregated model;
Optionally, the characteristics of image of meaning can be CNN feature herein.
Optionally, the Softmax layer in CNN can be set in Softmax disaggregated model;Correspondingly, head portrait attribute determines mould
Block 300, for the head portrait subject classification model according to pre-training, determining and described image feature is correspondingly, the network head portrait
In the probability distribution of each head portrait subject classification, can specifically include:
The Softmax layer that described image feature is imported to CNN, according to Softmax layers described, determining and described image feature
Correspondingly, probability distribution of the network head portrait in each head portrait subject classification.
Optionally, Figure 14 shows another structural block diagram of reference point determining device provided in an embodiment of the present invention, in conjunction with
Shown in Figure 12 and Figure 14, which can also include:
Reference sub-module constructs module 600, for determining positive sample user and negative sample user;Obtain each positive sample user
Behavioural characteristic and network head portrait head portrait attribute, as positive sample feature;And obtain the behavioural characteristic of each negative sample user
And the head portrait attribute of network head portrait, as negative sample feature;According to the machine learning algorithm training positive sample feature and bear
Sample characteristics obtain reference sub-model.
Above-described reference divides the function of determining device, can be realized by the program of server;The program can fill
It is loaded in the memory of server, and is called and implemented by the processor of server;Optionally, the hardware configuration of server can be as
Shown in Fig. 2, including:At least one processor and at least one processor;
The memory is stored with program, and the processor calls the program of the memory storage, and described program is used for:
Obtain the network head portrait of target user;
Determine the characteristics of image of the network head portrait;
The head portrait attribute of the network head portrait is determined according to described image feature;
According to the reference sub-model of pre-training, determination is corresponding to the behavioural characteristic of the head portrait attribute and the target user
Reference point;Wherein, behavioural characteristic and network head of the reference sub-model according at least to positive sample user and negative sample user
The head portrait attribute training of picture obtains, and the creditworthiness of positive sample user is higher than negative sample user.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.For device disclosed in embodiment
For, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place is said referring to method part
It is bright.
Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure
And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These
Function is implemented in hardware or software actually, the specific application and design constraint depending on technical solution.Profession
Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered
Think beyond the scope of this invention.
The step of method described in conjunction with the examples disclosed in this document or algorithm, can directly be held with hardware, processor
The combination of capable software module or the two is implemented.Software module can be placed in random access memory (RAM), memory, read-only deposit
Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology
In any other form of storage medium well known in field.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments in the case where not departing from core of the invention thought or scope.Therefore, originally
Invention is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein
Consistent widest scope.
Claims (12)
1. a kind of reference point determines method, which is characterized in that including:
Obtain the network head portrait of target user;
Determine the characteristics of image of the network head portrait;
The head portrait attribute of the network head portrait is determined according to described image feature;
According to the reference sub-model of pre-training, levy corresponding with the behavioural characteristic of the head portrait attribute and the target user is determined
Letter point;Wherein, behavioural characteristic and network head portrait of the reference sub-model according at least to positive sample user and negative sample user
The training of head portrait attribute obtains, and the creditworthiness of positive sample user is higher than negative sample user.
2. reference according to claim 1 point determines method, which is characterized in that described image feature is convolutional neural networks
The characteristics of image of CNN feature, the determination network head portrait includes:
Process of convolution is carried out to network head portrait by multiple convolutional layers of CNN, and the multiple convolution is connected by full articulamentum
Layer process of convolution as a result, formed setting dimension feature vector, obtain the CNN feature of network head portrait.
3. reference according to claim 2 point determines method, which is characterized in that multiple convolutional layers pair by CNN
Network head portrait carries out process of convolution:
The first layer convolutional layer that network head portrait is imported to CNN carries out process of convolution to network head portrait by first layer convolutional layer;
The convolution processing result of first layer convolutional layer is imported into next layer of convolutional layer and carries out process of convolution, and makes next layer of convolution
The convolution processing result of one layer of convolutional layer in layer process of convolution, until all convolutional layers carried out process of convolution;
The result for connecting the multiple convolutional layer process of convolution by full articulamentum includes:
The convolution processing result of final layer convolutional layer is imported into full articulamentum, connection forms the feature vector of setting dimension, obtains
The CNN feature of network head portrait.
4. reference according to claim 3 point determines method, which is characterized in that the convolution processing result of next layer of convolutional layer
Corresponding characteristic dimension, lower than the corresponding characteristic dimension of convolution processing result of upper one layer of convolutional layer.
5. reference according to claim 2 point determines method, which is characterized in that described to determine institute according to described image feature
The head portrait attribute for stating network head portrait includes:
According to the head portrait subject classification model of pre-training, determination is with described image feature correspondingly, the network head portrait is in each head
As the probability distribution of subject classification;Image corresponding figure of the head portrait subject classification model according at least to each head portrait subject classification
As feature training obtains.
6. reference according to claim 5 point determines method, which is characterized in that further include:
Obtain the image of each head portrait subject classification;
Obtain the corresponding characteristics of image of image of each head portrait subject classification;
According to Softmax sorting algorithm, the characteristics of image of the image of each head portrait subject classification of training obtains Softmax classification
Model.
7. reference according to claim 6 point determines method, which is characterized in that the Softmax disaggregated model is set to
The Softmax layer of CNN;The head portrait subject classification model according to pre-training is determined with described image feature correspondingly, described
Network head portrait includes in the probability distribution of each head portrait subject classification:
The Softmax layer that described image feature is imported to CNN determines corresponding to described image feature according to Softmax layers described
, probability distribution of the network head portrait in each head portrait subject classification.
8. reference according to claim 1 point determines method, which is characterized in that further include:
Determine positive sample user and negative sample user;
The behavioural characteristic of each positive sample user and the head portrait attribute of network head portrait are obtained, as positive sample feature;And it obtains each
The behavioural characteristic of negative sample user and the head portrait attribute of network head portrait, as negative sample feature;
According to the machine learning algorithm training positive sample feature and negative sample feature, reference sub-model is obtained.
9. a kind of reference divides determining device, which is characterized in that including:
Head portrait obtains module, for obtaining the network head portrait of target user;
Characteristics of image determining module, for determining the characteristics of image of the network head portrait;
Head portrait attribute determination module, for determining the head portrait attribute of the network head portrait according to described image feature;
Reference divides determining module, determining to use with the head portrait attribute and the target for the reference sub-model according to pre-training
The corresponding reference of the behavioural characteristic at family point;Wherein, the reference sub-model is according at least to positive sample user and negative sample user
The training of the head portrait attribute of behavioural characteristic and network head portrait obtains, and the creditworthiness of positive sample user is higher than negative sample user.
10. reference according to claim 9 divides determining device, which is characterized in that described image feature can be special for CNN
Sign, described image characteristic determination module are specifically included for determining the characteristics of image of the network head portrait:
Process of convolution is carried out to network head portrait by multiple convolutional layers of CNN, and the multiple convolution is connected by full articulamentum
Layer process of convolution as a result, formed setting dimension feature vector, obtain the CNN feature of network head portrait.
11. reference according to claim 10 divides determining device, which is characterized in that the head portrait attribute determination module is used
In the head portrait attribute for determining the network head portrait according to described image feature, specifically include:
According to the head portrait subject classification model of pre-training, determination is with described image feature correspondingly, the network head portrait is in each head
As the probability distribution of subject classification;Image corresponding figure of the head portrait subject classification model according at least to each head portrait subject classification
As feature training obtains.
12. a kind of server, which is characterized in that including:At least one processor and at least one processor;The memory is deposited
Program is contained, the processor calls the program of the memory storage, and described program is used for:
Obtain the network head portrait of target user;
Determine the characteristics of image of the network head portrait;
The head portrait attribute of the network head portrait is determined according to described image feature;
According to the reference sub-model of pre-training, levy corresponding with the behavioural characteristic of the head portrait attribute and the target user is determined
Letter point;Wherein, behavioural characteristic and network head portrait of the reference sub-model according at least to positive sample user and negative sample user
The training of head portrait attribute obtains, and the creditworthiness of positive sample user is higher than negative sample user.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710330665.2A CN108876586A (en) | 2017-05-11 | 2017-05-11 | A kind of reference point determines method, apparatus and server |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710330665.2A CN108876586A (en) | 2017-05-11 | 2017-05-11 | A kind of reference point determines method, apparatus and server |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108876586A true CN108876586A (en) | 2018-11-23 |
Family
ID=64319703
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710330665.2A Pending CN108876586A (en) | 2017-05-11 | 2017-05-11 | A kind of reference point determines method, apparatus and server |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108876586A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109448737A (en) * | 2018-08-30 | 2019-03-08 | 百度在线网络技术(北京)有限公司 | Creation method, device, electronic equipment and the storage medium of virtual image |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150254766A1 (en) * | 2014-03-05 | 2015-09-10 | Marc Abramowitz | System and method for generating a dynamic credit risk rating for a debt security |
CN105335712A (en) * | 2015-10-26 | 2016-02-17 | 小米科技有限责任公司 | Image recognition method, device and terminal |
CN105589798A (en) * | 2015-12-10 | 2016-05-18 | 小米科技有限责任公司 | Credit value calculation method and apparatus |
CN106447490A (en) * | 2016-09-26 | 2017-02-22 | 广州速鸿信息科技有限公司 | Credit investigation application method based on user figures |
CN106447625A (en) * | 2016-09-05 | 2017-02-22 | 北京中科奥森数据科技有限公司 | Facial image series-based attribute identification method and device |
-
2017
- 2017-05-11 CN CN201710330665.2A patent/CN108876586A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150254766A1 (en) * | 2014-03-05 | 2015-09-10 | Marc Abramowitz | System and method for generating a dynamic credit risk rating for a debt security |
CN105335712A (en) * | 2015-10-26 | 2016-02-17 | 小米科技有限责任公司 | Image recognition method, device and terminal |
CN105589798A (en) * | 2015-12-10 | 2016-05-18 | 小米科技有限责任公司 | Credit value calculation method and apparatus |
CN106447625A (en) * | 2016-09-05 | 2017-02-22 | 北京中科奥森数据科技有限公司 | Facial image series-based attribute identification method and device |
CN106447490A (en) * | 2016-09-26 | 2017-02-22 | 广州速鸿信息科技有限公司 | Credit investigation application method based on user figures |
Non-Patent Citations (1)
Title |
---|
卢九评编著: "微信营销技巧及案例", 《微信营销技巧及案例》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109448737A (en) * | 2018-08-30 | 2019-03-08 | 百度在线网络技术(北京)有限公司 | Creation method, device, electronic equipment and the storage medium of virtual image |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107239802B (en) | Image classification method and device | |
CN108073659A (en) | A kind of love and marriage object recommendation method and device | |
CN109816092A (en) | Deep neural network training method, device, electronic equipment and storage medium | |
CN107423442A (en) | Method and system, storage medium and computer equipment are recommended in application based on user's portrait behavioural analysis | |
CN110533097A (en) | A kind of image definition recognition methods, device, electronic equipment and storage medium | |
CN107808132A (en) | A kind of scene image classification method for merging topic model | |
CN110348428A (en) | Eye fundus image classification method, device and computer readable storage medium | |
CN107194898A (en) | The method for pushing of the methods of exhibiting of hotel's image, storage medium and hotel information | |
CN107742107A (en) | Facial image sorting technique, device and server | |
CN108229269A (en) | Method for detecting human face, device and electronic equipment | |
CN112231584B (en) | Data pushing method and device based on small sample transfer learning and computer equipment | |
CN108681970A (en) | Finance product method for pushing, system and computer storage media based on big data | |
CN108229523A (en) | Image detection, neural network training method, device and electronic equipment | |
JP2018055470A (en) | Facial expression recognition method, facial expression recognition apparatus, computer program, and advertisement management system | |
CN107784372A (en) | Forecasting Methodology, the device and system of destination object attribute | |
CN107785061A (en) | Autism-spectrum disorder with children mood ability interfering system | |
CN110196945B (en) | Microblog user age prediction method based on LSTM and LeNet fusion | |
CN110097090A (en) | A kind of image fine granularity recognition methods based on multi-scale feature fusion | |
CN107609563A (en) | Picture semantic describes method and device | |
CN109919006A (en) | Expression detection method, device, electronic equipment and storage medium | |
CN110163250A (en) | Image desensitization process system, method and device based on distributed scheduling | |
CN108876193A (en) | A kind of air control model building method based on credit score | |
CN107862322A (en) | The method, apparatus and system of picture attribute classification are carried out with reference to picture and text | |
CN109670927A (en) | The method of adjustment and its device of credit line, equipment, storage medium | |
CN108776846A (en) | Recommend method, apparatus, computer equipment and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |