CN115511606A - Object identification method, device, equipment and storage medium - Google Patents
Object identification method, device, equipment and storage medium Download PDFInfo
- Publication number
- CN115511606A CN115511606A CN202211261970.8A CN202211261970A CN115511606A CN 115511606 A CN115511606 A CN 115511606A CN 202211261970 A CN202211261970 A CN 202211261970A CN 115511606 A CN115511606 A CN 115511606A
- Authority
- CN
- China
- Prior art keywords
- vector
- type
- node
- feature
- target
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Mathematical Analysis (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- General Engineering & Computer Science (AREA)
- Computational Mathematics (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Algebra (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Animal Behavior & Ethology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Operations Research (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The application provides an object identification method, device, equipment and storage medium, which are used for acquiring portrait characteristic data and incidence relation of an object to be identified so as to generate a map structure of the object to be identified, further utilizing a behavior type prediction model obtained by learning the map structure by using a side type parameter sharing neighborhood information fusion mechanism, determining characteristic behavior probability distribution of the object to be identified, and determining whether the object to be identified is a characteristic behavior object according to the characteristic behavior probability distribution. The behavior type prediction model obtained by adopting the edge type parameter sharing neighborhood information fusion mechanism is suitable for the map with a large number of nodes, reduces the number of neural networks required by model construction, reduces the calculation amount and scale of the model, avoids the problems of gradient disappearance and gradient explosion, and improves the learning efficiency and the prediction efficiency. The distribution difference between the neighborhood nodes is considered, so that the behavior type prediction model is fully learned, and the object to be recognized with the characteristic behavior risk is accurately and efficiently recognized.
Description
Technical Field
The present application relates to the field of financial technology (Fintech), and in particular, to an object recognition method, apparatus, device, and storage medium.
Background
With the rapid development of computer technology and internet technology, financial technology (Fintech) is becoming a hot point for the innovative development of financial industry as a product of the deep integration of finance and technology. For financial institutions, it is crucial to identify aspects such as security of the enterprise objects to identify characteristic behaviors, e.g., whether the objects are at risk of the characteristic behavior, such as fraud.
For the identification of whether an enterprise object has a characteristic behavioral risk, a deep learning model based on a knowledge graph is generally realized. However, in the prior art, when a learning model is constructed to calculate neighborhood information propagation, the relationship between the number of neural networks to be trained and the scale (number of nodes) of the knowledge graph is as follows: the number of the neural networks = O (| number of nodes |) 2 ). Therefore, in the prior art, too many neural networks need to be trained, and both memory resources and computational resources in the training process increase in a square exponential manner with the increase of the number of nodes, so that the model calculation amount is increased, the training efficiency is lowered, and the method is not suitable for the maps with larger number of nodes. In addition, when the excessive neural network is used for T times of iterative training, the problem of gradient disappearance or gradient explosion is easy to occur due to the large number of nodes, and further the mode is causedType training fails, reducing training and prediction efficiency.
In addition, in the use process of the existing deep learning model constructed based on the knowledge graph, the subgraphs divided by edges of different types have great variance no matter the number of the edges or the number of the nodes, and the variance is caused by the unbalanced distribution of the edges on the nodes. In the prior art, nodes are ordered, and then model training and iterative computation are completed by using a shared transfer matrix, so that more neural networks cannot be trained, and further, the trained model has poor prediction capability and inaccurate prediction results.
Disclosure of Invention
The application provides an object recognition method, device, equipment and storage medium, which are used for solving the technical problems of low model training efficiency, low prediction efficiency and poor prediction capability when a deep learning model is constructed based on a map for object recognition in the prior art.
In a first aspect, the present application provides an object recognition method, including:
acquiring portrait feature data and an association relation of an object to be identified, and generating a map structure of the object to be identified according to the portrait feature data and the association relation, wherein the object to be identified comprises an enterprise object or an individual object;
obtaining characteristic behavior probability distribution of the object to be identified according to the graph structure and a behavior type prediction model, wherein the behavior type prediction model is obtained by learning the graph structure by adopting a side type parameter sharing neighborhood information fusion mechanism;
and determining whether the object to be identified is a characteristic behavior object according to the characteristic behavior probability distribution, wherein the characteristic behavior object has a characteristic behavior risk.
In one possible design, learning the graph structure by using a parameter sharing neighborhood information fusion mechanism of the edge type to obtain the behavior type prediction model includes:
determining target feature data of a target node according to the graph structure, wherein the target feature data comprises feature data transmitted by each neighborhood node of the target node through edges of different types, and the target node is a node which characterizes the object to be identified in the graph structure;
determining new feature data of the target node during the target iteration times according to the feature data transmitted by different types of edges;
and obtaining a node vector according to the newly added feature data and the self-hidden layer vector of the target node in the previous target iteration time, wherein the self-hidden layer vector is used for representing the feature data of the target node self-fused in the previous target iteration time, and the node vector is used for representing the behavior type prediction model.
In one possible design, the determining target feature data of the target node according to the graph structure includes:
determining a weight coefficient of each type of edge of the target node in the target iteration times according to the graph structure, wherein the weight coefficient is used for representing the correlation between the target node and each neighborhood node;
determining feature data transmitted by the edges of different types according to the weight coefficient corresponding to each type of edge and a weight parameter network model;
the weight parameter network model comprises a first feature mapping function and a feature matrix, wherein the first feature mapping function is related to the target iteration times and used for feature space conversion, and the feature matrix is related to the edge type of the target node and the target iteration times.
In one possible design, the determining, according to the graph structure, a weight coefficient of each type of edge of the target node at the target iteration number includes:
obtaining a weight coefficient corresponding to each type of edge of the target node through differentiation fusion processing according to the graph structure and the feature mapping function cluster;
the sum of the weight coefficients corresponding to the same type of edges of the target node is 1, and the feature mapping function cluster comprises a set of second feature mapping functions corresponding to each type of edges;
in a possible design, the obtaining, by differential fusion processing according to the graph structure and the feature mapping function set, a weight coefficient corresponding to each type of edge of the target node includes:
for each type of edge of the target node, performing matrix multiplication operation according to a node vector corresponding to the target node and each neighborhood node in the previous time of the target iteration times and a set of second feature mapping functions corresponding to the current type of edge of the target node to obtain a matrix multiplication operation result;
carrying out equal-scale scaling processing on the matrix multiplication result to obtain a scaling processing result;
and normalizing the scaling processing result by utilizing a normalization index function to obtain a weight coefficient corresponding to each type of edge of the target node.
In one possible design, the determining feature data passed by the different types of edges according to the weight coefficient corresponding to each type of edge and a weight parameter network model includes:
and obtaining a product between the weight coefficient corresponding to each type of edge, the node vector of the neighborhood node corresponding to each type of edge in the previous time of the target iteration times and the weight parameter network model corresponding to each type of edge, and determining the obtained product as feature data transmitted by the edge of the corresponding type.
In one possible design, the determining, according to feature data delivered by each of the different types of edges, new feature data of the target node at the target iteration time includes:
acquiring the sum of the feature vectors corresponding to different types of edges of the target node to obtain the sum of the vectors, wherein the feature vector corresponding to each type of edge is used for representing feature data transmitted by each type of edge;
performing piecewise linear processing on the vector sum by adopting an activation function to obtain a linear result vector;
and acquiring the product of the linear mapping matrix and the linear result vector, and representing the newly added feature data of the target node when the target iteration times is carried out by using the acquired product vector.
In a possible design, the obtaining a node vector according to the newly added feature data and a self hidden layer vector of the target node at a previous time of the target iteration number includes:
and inputting the product vector and the self hidden layer vector of the target node when the target iteration times is the previous time into a preset recurrent neural network, and determining the output as the node vector.
In one possible design, obtaining the probability distribution of the object to be recognized according to the graph structure and behavior type prediction model includes:
mapping the node vector into a category probability distribution vector through a third feature mapping function;
and inputting the category probability distribution vector into a two-classifier to obtain a category output vector, wherein the category output vector is used for representing the characteristic behavior probability distribution of the object to be identified.
In one possible design, the determining whether the object to be identified is a feature behavior object according to the feature behavior probability distribution includes:
and judging whether the object to be identified is the characteristic behavior object or not according to the category output vector and a preset category vector, wherein the preset category vector comprises a category vector of the characteristic behavior object and a category vector which is not the characteristic behavior object.
In one possible design, obtaining the association relationship of the object to be recognized includes:
acquiring an associated object having a transaction with the object to be recognized according to the portrait characteristic data of the object to be recognized;
determining the relationship category between each associated object and the object to be identified according to the transaction data corresponding to the transaction;
and obtaining the incidence relation of the object to be identified according to the relation category.
In one possible design, the generating a graph structure of the object to be recognized according to the portrait feature data and the association relation includes:
setting the object to be identified as a node of the graph structure, and setting the edge of the graph structure according to the incidence relation of the object to be identified so as to generate the graph structure of the object to be identified.
In a possible design, after the obtaining of the portrait characteristic data of the object to be recognized, the method further includes:
and performing feature value on the image feature data of the object to be recognized, and determining the image feature data after feature value is performed as an initial vector of the target node in the graph structure.
In a second aspect, the present application provides an object recognition apparatus, comprising:
the system comprises a first processing module, a second processing module and a third processing module, wherein the first processing module is used for acquiring portrait characteristic data and an association relation of an object to be identified, and generating a map structure of the object to be identified according to the portrait characteristic data and the association relation, and the object to be identified comprises an enterprise object or an individual object;
the second processing module is used for obtaining the characteristic behavior probability distribution of the object to be recognized according to the map structure and a behavior type prediction model, wherein the behavior type prediction model is obtained by learning the map structure by adopting a side-type parameter sharing neighborhood information fusion mechanism;
and the third processing module is used for determining whether the object to be identified is a characteristic behavior object according to the characteristic behavior probability distribution, wherein the characteristic behavior object has a characteristic behavior risk.
In a third aspect, the present application provides an electronic device, comprising: a processor, and a memory communicatively coupled to the processor;
the memory stores computer-executable instructions;
the processor executes computer-executable instructions stored by the memory to implement any one of the possible object recognition methods provided in the first aspect.
In a fourth aspect, the present application provides a computer-readable storage medium having stored therein computer-executable instructions for implementing any one of the possible object recognition methods provided in the first aspect when executed by a processor.
In a fifth aspect, the present application provides a computer program product comprising computer executable instructions for implementing any one of the possible object recognition methods provided in the first aspect when executed by a processor.
The application provides an object recognition method, device, equipment and storage medium, which comprises the steps of firstly obtaining portrait characteristic data and incidence relation of an object to be recognized, further generating a map structure of the object to be recognized according to the portrait characteristic data and the incidence relation, and then obtaining characteristic behavior probability distribution of the object to be recognized according to the map structure and a behavior type prediction model, wherein the behavior type prediction model is obtained by adopting a side type parameter sharing neighborhood information fusion mechanism and map structure training. And determining whether the object to be identified is the characteristic behavior object according to the characteristic behavior probability distribution. The behavior type prediction model is constructed by adopting an edge type parameter sharing neighborhood information fusion mechanism, model parameters cannot increase exponentially along with the increase of the number of nodes, so that the method is suitable for the map with larger number of nodes, the number of neural networks required by model learning is reduced, the calculated amount and scale of the model can be effectively reduced, the problems of gradient disappearance and gradient explosion are avoided, and the learning efficiency and the prediction efficiency are improved. In addition, the parameter sharing neighborhood information fusion mechanism of the edge type increases the distribution difference among the neighborhood nodes through differential fusion, so that the constructed behavior type prediction model can be fully learned, and the object to be recognized with the characteristic behavior risk can be accurately and efficiently recognized.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to these drawings without inventive exercise.
Fig. 1 is a schematic view of an application scenario provided in an embodiment of the present application;
fig. 2 is a schematic flowchart of an object identification method according to an embodiment of the present application;
fig. 3 is a schematic flowchart of another object identification method according to an embodiment of the present application;
FIG. 4 is a schematic illustration of a portion of one pattern structure provided by an embodiment of the present application;
fig. 5 is a schematic flowchart of another object identification method according to an embodiment of the present application;
fig. 6 is a schematic flowchart of another object identification method according to an embodiment of the present application;
fig. 7 is a schematic flowchart of another object identification method according to an embodiment of the present application;
fig. 8 is a schematic structural diagram of an object recognition apparatus according to an embodiment of the present disclosure;
fig. 9 is a schematic structural diagram of another object recognition apparatus according to an embodiment of the present application;
fig. 10 is a schematic structural diagram of an electronic device according to an embodiment of the present application.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of methods and apparatus consistent with certain aspects of the present application, as detailed in the appended claims.
The terms "first," "second," "third," "fourth," and the like in the description and in the claims of the present application and in the above-described drawings (if any) are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are, for example, capable of operation in sequences other than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
For identifying whether an enterprise object or an individual object has a characteristic behavior risk, a deep learning model based on a knowledge graph is generally realized. However, in the prior art, when a learning model is constructed to calculate neighborhood information propagation, the relationship between the number of neural networks to be trained and the scale (number of nodes) of the knowledge graph is as follows: the number of the neural networks = O (| number of nodes |) 2 ). Therefore, in the prior art, too many neural networks need to be trained, and both memory resources and computational resources in the training process increase in a square exponential manner with the increase of the number of nodes, so that the model calculation amount is increased, the training efficiency is lowered, and the method is not suitable for the maps with larger number of nodes. In addition, when the excessive neural networks are used for T times of iterative training, the problem of gradient disappearance or gradient explosion is easily caused due to the large number of nodes, so that the model training fails, and the training and prediction efficiency is reduced. In addition, in the use process of the existing deep learning model constructed based on the knowledge graph, the subgraphs divided by edges of different types have great variance no matter the number of the edges or the number of the nodes, and the variance is caused by the unbalanced distribution of the edges on the nodes. In the prior art, node sequencing is performed to complete model training and iterative computation by using a shared transfer matrix, so that more neural networks cannot be trained, the trained model has poor prediction capability, and the prediction result is inaccurate.
In view of the foregoing problems in the prior art, the present application provides an object identification method, apparatus, device and storage medium. The inventive concept of the object recognition method provided by the present application lies in: the method is characterized in that an edge type parameter sharing neighborhood information fusion mechanism is provided to construct a behavior type prediction model, different types of edges are distinguished when the propagation of neighborhood information is calculated, the information propagation on each neighborhood node of a target node, namely transmitted characteristic data, is calculated according to the type of the edge, so that an original transfer matrix mode of different parameters of the edge type is replaced by a neural network shared by the parameters on the edge type, the model parameters cannot exponentially increase along with the increase of the number of the nodes, the construction method of the behavior type prediction model is suitable for the graph with larger number of the nodes, the number of the neural networks required by model training can be reduced, the calculated amount and the training scale of the model can be effectively reduced, the problems of gradient disappearance and gradient explosion are avoided, and the training efficiency and the prediction efficiency are improved. In addition, the parameter sharing neighborhood information fusion mechanism of the edge type can feed back the actual condition that the edges of the graph are distributed unevenly on the nodes through the weight coefficients of the edges of different types, so that the distribution difference among the neighborhood nodes can be fully considered, the behavior type prediction model obtained by training can be fully trained, and the accuracy of the object recognition capability can be effectively improved.
An exemplary application scenario of the embodiments of the present application is described below.
Fig. 1 is a schematic view of an application scenario provided by an embodiment of the present application, and as shown in fig. 1, a network is used to provide a medium for a communication link between a first electronic device 100 and a second electronic device 200, and the network may include various connection types, such as a wired connection, a wireless communication link, or a fiber optic cable. The first electronic device 100 and the second electronic device 200 may interact with each other through a network to receive or transmit a message. The first electronic device 100 may be any terminal capable of acquiring data such as image feature data and association of an object to be recognized, and the first electronic device 100 may be disposed at the object to be recognized itself, or disposed at a third party institution, and the like. The second electronic device 200 may be deployed at a financial institution or the like and configured to execute the object recognition method provided in the embodiment of the present application, and the second electronic device 200 acquires the portrait feature data and the association relation of the object to be recognized from the first electronic device 100, and then executes the object recognition method provided in the embodiment of the present application to recognize whether the object to be recognized is a feature behavior object.
Alternatively, the object to be identified may include each business object, organization group, or individual object, etc., and the characteristic behavior object refers to a business object, organization group, or individual object, etc. with a characteristic behavior risk, which may be, for example, fraud risk, etc.
It should be noted that, the embodiment of the present application does not limit the type of the first electronic device 100 described above, for example, the first electronic device 100 may be a computer, a smart phone, smart glasses, a smart bracelet, a smart watch, a tablet computer, and the like, and the first electronic device 100 in fig. 1 is illustrated as a computer. And the second electronic device 200 may be a server, a computer, a server cluster, etc., and is illustrated as a server in fig. 1.
It should be noted that the above application scenarios are only exemplary, and the object identification method provided in the embodiment of the present application includes, but is not limited to, the above application scenarios.
Fig. 2 is a schematic flowchart of an object identification method according to an embodiment of the present application. As shown in fig. 2, an object identification method provided in an embodiment of the present application includes:
s101: and acquiring portrait feature data and an association relation of the object to be recognized, and generating a map structure of the object to be recognized according to the portrait feature data and the association relation.
The method comprises the steps of obtaining data for representing independent attributes of an object to be recognized, such as portrait characteristic data of the object to be recognized, and obtaining data for representing hidden relations between the object to be recognized and other objects, such as incidence relations of the object to be recognized. And then constructing the atlas structure of the object to be identified according to the portrait feature data and the incidence relation of the object to be identified.
Taking the object to be recognized as an enterprise object as an example, the image feature data of the object to be recognized may be, for example, enterprise basic information such as an industry type, enterprise legal person information, an enterprise registration address, an enterprise legal person type, an enterprise technical number, enterprise business category information such as an enterprise registration age, an enterprise business information change frequency, an enterprise legal person information change frequency, enterprise business category information such as an enterprise loan/deposit balance, a deposit monthly amount, enterprise deposit information such as an enterprise loan/deposit balance, a deposit monthly amount, a half (full) year transfer frequency, a half (full) year transfer amount, and credit investigation information such as an enterprise blacklist and an enterprise credit investigation. The specific content of the portrait feature of the object to be recognized is determined by the type of the object to be recognized, which is not limited in the embodiment of the present application.
In a possible design, a possible implementation manner of acquiring the association relationship of the object to be identified in step S101 includes:
acquiring associated objects with the objects to be identified, which have transaction traffic with the objects to be identified, based on the image characteristic data of the objects to be identified, further determining the relationship types between the associated objects and the objects to be identified according to the transaction data corresponding to the transaction traffic, and then determining the association relationship of the objects to be identified according to the relationship types, wherein one relationship type corresponds to one association relationship.
For example, the object to be recognized is an enterprise object, and the determined relationship category may include, but is not limited to, the following:
first, an upstream and downstream transfer relationship occurs for inter-enterprise transfers; second, sharing important affiliates between enterprises, such as corporate relationships, guarantor relationships, affiliate relationships, etc.; thirdly, the trade relationship of the trade transaction frequency, the amount, the single amount, the transaction times, the transaction time range, the transaction place and the like of the two parties; fourthly, the equity relationship of mutual stock control among enterprises; fifth, for example, shared portrait characteristics data such as registration address, telephone number, emergency contact, registration address, bank account, company mailbox, etc.; sixth, customer relationships with similar customer group relationships.
In one possible design, the generating a graph structure of the object to be recognized according to the portrait feature data and the association relationship in step S101 includes:
setting the object to be identified as a node of the graph structure, setting an edge of the graph structure according to the incidence relation of the object to be identified, namely, expressing one edge by one incidence relation represented by one relation type, thereby constructing the graph structure of the object to be identified. After the portrait feature data of the object to be recognized is obtained, the portrait feature data of the object to be recognized is also subjected to feature value, and the portrait feature data subjected to feature value is determined as an initial vector of a target node in a map structure and can be expressed as an initial vector of a target node in the map structureThat is, the target node has not yet started the iteration of the target iteration number. v denotes a target node, i.e. a node in the graph structure representing an object to be identified.
It should be noted that, the initial vector has dimensions of 1 × M, and a feature space with a higher dimension may be used for better classification effect during training, for example, the dimension of the target node may be set to 300 dimensions. And if the number of the portrait feature data is less than the dimension of the initial vector, the value of the dimension can adopt 0 place.
S102: and obtaining the characteristic behavior probability distribution of the object to be recognized according to the map structure and the behavior type prediction model.
The behavior type prediction model is obtained by learning the map structure by adopting an edge type parameter sharing neighborhood information fusion mechanism.
After the map structure of the object to be recognized is constructed, learning is carried out by utilizing the map structure and the parameter sharing neighborhood information fusion mechanism of the edge type to obtain the behavior type prediction model, in other words, the behavior type prediction model is constructed by adopting the parameter sharing neighborhood information fusion mechanism of the edge type and the map result through training. Thus, the behavior type prediction model is a feature expression of the object to be recognized. And then obtaining the characteristic behavior probability distribution of the object to be recognized through the behavior type prediction model and the two classifiers, and realizing whether the object to be recognized is the characteristic behavior object.
In the process of obtaining the behavior type prediction model by learning the graph structure by adopting the edge type parameter sharing neighborhood information fusion mechanism, feature data transmitted by different types of edges can be considered, so that each neighborhood information of the target node can be differentially transmitted. In addition, the feature data transmitted by each neighborhood node is obtained based on different types of edges, the edges of different types can distinguish different incidence relations of the object to be recognized, the different incidence relations are different feature extraction networks in the model building process, and equivalently, a mode of adopting a neural network with shared parameters on the edge types to replace a transfer matrix with the edge types without sharing the parameters in the prior art is adopted, so that the model parameters in the model building process cannot be exponentially increased along with the increase of the number of the nodes, the network parameter sharing on the edge types can also reduce the number of the neural networks required in the model building process, the calculated amount and the training scale of the parameters in the model building process are reduced, the model convergence speed is higher, the gradient disappearance problem and the gradient explosion problem can also be avoided, and the model training efficiency and the prediction efficiency are improved.
In one possible design, a possible implementation manner of learning the graph structure by using the edge-type parameter-sharing neighborhood information fusion mechanism in step S102 to obtain the behavior type prediction model is shown in fig. 3. Fig. 3 is a schematic flowchart of another object identification method according to an embodiment of the present application. As shown in fig. 3, the embodiment of the present application includes:
s201: and determining target characteristic data of the target node according to the map structure.
The target feature data comprises feature data transmitted by each neighborhood node of a target node through edges of different types, and the target node is a node representing an object to be identified in a graph structure.
Based on the graph structure, feature data transmitted by each neighborhood node of the target node through different types of edges in the graph structure is obtained, namely the target feature data of the target node is determined according to the graph structure.
For example, the neighborhood structure of the target node v in the graph structure is shown in fig. 4, the target node v has three neighborhood nodes p, w, and s, and each node has an association relationship: r 1(v,p) 、R 1(v,s) And R 5(v,w) 、R 5(v,s) . Wherein R is 1(v,p) Representing the object to be identified characterized by the target node v and the associated object p (characterized by the neighborhood nodes) having transaction with the object to be identified, wherein the relationship class is R 1 By analogy to R 1(v,s) And R 5(v,w) R 5(v,s) 。
Thus, this step calculates each neighborhood node { p, s, w } of the target node based on the graph results by different edge types, e.g., { R } 1 ,R 5 Characteristic data propagated on }Andi.e. target characteristic data of the target nodeAndin the process of constructing the model, fixed iteration is carried out for multiple times, and each iteration is called information propagation at a moment t, namely feature data transmission. After multiple iterative computations, the feature expression of the target node is regarded as the node vector of the target node after learning the graph structure.
S202: and determining the newly added feature data of the target node during the target iteration times according to the feature data transmitted by the edges of different types.
After obtaining target feature data of the target node, namely feature data transmitted by different types of edges of the target node, determining new feature data of the target node during target iteration times, namely new feature data transmitted by different types of edges based on the feature data transmitted by different types of edges of the target nodeAt time t the atlas structure is populated by propagating new information, e.g. toIndicating that at time t, the target node v is being defined by different types of edges (e.g., R) 1(v,p) 、R 1(v,s) And R 5(v,w) 、R 5(v,s) ) And the newly added neighborhood information on the composition formed by the same composition is summarized. In other words, the step is carried out byAndcalculating out
S203: and obtaining a node vector according to the newly added feature data and the self hidden layer vector of the target node when the target iteration times is the previous time.
The self-hidden layer vector is used for representing feature data of the target node, which are self-fused in the previous time of the target iteration times, and the node vector is used for representing the behavior type prediction model.
For the target node, at the time t, the target node hides the layerThe vector, namely the feature data to be expressed by the node vector comes from two parts, one part is newly added feature data, namely the feature dataThe other part comes from the hidden layer vector coding of the target node at the previous moment (namely t-1) and the target node isTherefore, the step is realized by adding new feature dataAnd self-hidden vectorCalculating a hidden layer vector of a target node at time tI.e. a node vector. And the calculated node vector represents a behavior type prediction model.
The object identification method provided by the embodiment of the application comprises the steps of firstly obtaining feature data transmitted by each neighborhood node through different types of edges, further obtaining newly added feature data of a target node according to the feature data, then obtaining a node vector for expressing a behavior type prediction model by combining a self-hidden vector of the target node when the target node is previous to the target iteration times, finishing the learning of a graph structure, and obtaining the behavior type prediction model for predicting whether an object to be identified is a feature behavior object.
On the basis of obtaining the behavior type prediction model, the implementation manner of obtaining the probability distribution possibility of the object to be recognized according to the graph structure and the behavior type prediction model in step S102 includes:
firstly, the node vectors are mapped into category probability distribution vectors through a third feature mapping function, then the category probability distribution vectors are input into a two-classifier to obtain category output vectors, and the category output vectors are used for representing feature behavior probability distribution of the object to be identified.
In order to predict whether the object to be recognized is a feature behavior object, a classifier may be used to classify the feature expression of the object to be recognized. For example, the characteristic behavior probability distribution is obtained by using the following formula (1):
wherein, the first and the second end of the pipe are connected with each other,representing class output vector, sigmoidThe function represents a two-classifier that is,representing a class probability distribution vector.
In particular, the Sigmoid function is used for the implementation of the two-classifier, since the bernoulli distribution of the Sigmoid function over the probability q and the probability (1-q) is very efficient, with an output range of (0,1). In order to make the node vectorBeing able to input to the classifier, it is necessary to feature the dimensions of the node vector, e.g. 1 × M using a third feature mapping function characterized by SThe vectors are mapped into 1 x 2 class probability distribution vectors.
For example, the mapping process by the third feature mapping function is as shown in the following equation (2):
wherein the content of the first and second substances,it is a characteristic expression f of M dimension; the third feature mapping function S' is a vector of dimension mx 2;representing a class probability distribution vector.
The Sigmoid function is calculated as shown in equation (3):
from the above equations (1) to (3), the following equation (4) can be obtained:
the meaning expressed by the characteristic behavior probability distribution of the class output vector characterization is shown in the following formula (5):
it should be noted that the two classifiers used when the two classifiers are used to obtain the class output vector are trained two classifiers, that is, before the class probability distribution vector is input to the two classifiers, the two classifiers need to be trained by using the training sample, so that the trained two classifiers can identify the object to be identified as the characteristic behavior object or the non-characteristic behavior object.
Optionally, the two classifiers are trained using a loss function as shown in equation (6) below:
wherein L represents a loss function, D l Representing sample objects comprising characteristic behavior sample objects and non-characteristic behavior sample objects, y' representing that the object represented by the node u is a label of the characteristic behavior object,and representing that the object represented by the two-classifier prediction node u is a label of the characteristic behavior object.
As can be seen from the formula (6), the maximum likelihood estimates of all nodes with characteristic behavior objects are accumulated to obtain the loss value of the two-classifier, the two-classifier is trained by using the loss value, and parameters are adjusted during training to fit data characteristics. Wherein the loss function is a cross entropy function.
S103: and determining whether the object to be identified is the characteristic behavior object or not according to the characteristic behavior probability distribution.
After the behavior probability distribution of the object to be recognized is obtained, whether the object to be recognized is a characteristic behavior object is determined based on the behavior probability distribution of the object to be recognized, wherein the characteristic behavior object has characteristic behavior risks, and therefore whether the object to be recognized has the characteristic behavior risks is recognized.
For example, the category output vector representing the characteristic behavior probability distribution is compared with a preset category vector, and whether the object to be identified is the characteristic behavior object is judged. The preset category vector comprises a category vector which is a characteristic behavior object and a category vector which is a non-characteristic behavior object.
For example, the class vector for the feature behavior object is [1,0], and the class vector for the non-feature behavior object is [0,1]. If the category output vector is the former, it indicates that the object to be recognized is recognized as a characteristic behavior object, that is, the object to be recognized has a characteristic behavior risk. If the category output vector is the latter, it indicates that the object to be recognized is recognized as a non-characteristic behavior object, that is, the object to be recognized has no characteristic behavior risk.
It can be understood that the preset class vector is obtained by semantic labeling when training the two classifiers, and the purpose of the preset class vector is to interpret the output semantics of the two classifiers. For example, the labeling method is as follows:
if the characteristic behavior object is found, the category is 1, and the corresponding category vector is [1,0]; if the behavior object is not a feature behavior object, the class vector is [0,1]. The labeling mode is not limited in the embodiment of the application, and can be set according to a specific recognition task.
Therefore, whether the object to be recognized is the characteristic behavior object can be efficiently and accurately recognized through the object recognition method provided by the embodiment of the application. Table 1 shows the accuracy of constructing a behavior type prediction model and the accuracy of identifying an object to be identified by using the object identification methods provided in the prior art and the embodiments of the present application, respectively.
TABLE 1
As can be seen from table 1, the model construction accuracy and the recognition accuracy of the object recognition method provided in the embodiment of the present application are higher than those of the prior art, and the difference between the two is obvious for the recognition accuracy, which indicates that the behavior type prediction model constructed in the object recognition method provided in the embodiment of the present application has a stronger generalization capability, specifically, the behavior type prediction model is trained and learned more sufficiently, and the behavior type prediction model shares parameters on the edge types, so that the model parameters can be effectively adjusted by the sufficient training and learning, which is helpful for improving the prediction capability. And the differentiation between different neighborhood information can be pulled open by a differentiated neighborhood fusion mode, so that the effect of expanding the differentiation between the neighborhood information is achieved, and the generalization capability of the behavior type prediction model is improved.
Table 2 shows the calculation resources and memory resources consumed when the behavior type prediction model is constructed for iterative computation by using the object identification method provided in the embodiments of the present application and the prior art.
TABLE 2
As can be seen from table 2, in the prior art scheme, a large amount of CPU resources are consumed due to the existence of a large amount of redundant parameters, and the number of parameters is too large, so that a large amount of memory resources are required for storing variables and intermediate values.
The object recognition method provided by the embodiment of the application comprises the steps of firstly obtaining the portrait characteristic data and the incidence relation of an object to be recognized, further generating the atlas structure of the object to be recognized according to the portrait characteristic data and the incidence relation, and then obtaining the characteristic behavior probability distribution of the object to be identified according to the atlas structure and the behavior type prediction model, wherein the behavior type prediction model is obtained by adopting an edge type parameter sharing neighborhood information fusion mechanism and atlas structure training. And determining whether the object to be identified is the characteristic behavior object according to the characteristic behavior probability distribution. The behavior type prediction model is constructed by adopting an edge type parameter sharing neighborhood information fusion mechanism, model parameters cannot increase exponentially along with the increase of the number of nodes, so that the method is suitable for the map with the larger number of nodes, the number of neural networks required by model training is reduced, the calculated amount and the training scale of the model can be effectively reduced, the problems of gradient disappearance and gradient explosion are avoided, and the training efficiency and the prediction efficiency are improved. In addition, the parameter sharing neighborhood information fusion mechanism of the edge type increases the distribution difference among the neighborhood nodes through differential fusion, so that the constructed behavior type prediction model can be fully trained, and the object to be recognized with the characteristic behavior risk can be accurately and efficiently recognized.
In one possible design, a possible implementation of step S201 is shown in fig. 5. Fig. 5 is a schematic flowchart of another object identification method according to an embodiment of the present application. As shown in fig. 5, the embodiment of the present application includes:
s301: and determining the weight coefficient of each type of edge of the target node in the target iteration times according to the graph structure.
The weight coefficient is used for representing the correlation between the target node and each neighborhood node.
And determining the weight coefficient of each type of edge of the target node at the target iteration number based on the graph structure. Taking fig. 4 as an example, namely, determining each type of edge (R) of the target node v based on the graph structure shown in fig. 4 1(v,p) 、R 1(v,s) 、R 5(v,w) 、R 5(v,s) ) Weighting factor at target number of iterations (t) Indicates that the edge type is R 1 The relevance of the neighborhood nodes p and v at time t, and so on.
The weight coefficient for representing the correlation is related to the target iteration times, namely the fixed iteration times and the edge types, and different edge types and the target iteration times correspond to different weight coefficients.
In one possible design, possible implementations of step S301 include:
and obtaining a weight coefficient corresponding to each type of edge of the target node through differentiation fusion processing according to the map structure and the feature mapping function cluster. Namely, difference melting processing is carried out through the feature mapping function cluster based on the graph structure, and the weight coefficient corresponding to each type of edge of the target node is obtained.
The feature mapping function cluster comprises a set of second feature mapping functions corresponding to each type of edge, the feature mapping function cluster is used for feature dimension conversion, each feature mapping layer is a neural network and has a trainable parameter set, the parameter set is related to the target iteration times and the edge type, and different edges have different second feature mapping functions.
It should be noted that performing feature dimension conversion specifically refers to converting a low dimension into a high dimension. The high-dimensional space can better orthogonalize the low-dimensional features, the orthogonalized features are mutually independent, and the main components of the features can be better reflected, so that the distinguishing degree between the data features is larger, and the essence of the differential fusion processing is.
In addition, for edges of the same type, the sum of their corresponding weighting factors is 1, for exampleAndthe sum of the total number of the carbon atoms is 1,andthe sum is 1.
In a possible design, the weight coefficient corresponding to each type of edge of the target node is obtained through differentiation fusion processing according to the graph structure and the feature mapping function cluster, and possible implementation manners include matrix multiplication, equal scaling processing and normalization, and the implementation steps are shown in fig. 6. Fig. 6 is a schematic flowchart of another object identification method according to an embodiment of the present application. As shown in fig. 6, the embodiment of the present application includes:
s401: aiming at each type of edge of the target node, performing matrix multiplication operation according to the node vectors corresponding to the target node and each neighborhood node in the previous time of the target iteration times and a set of second feature mapping functions corresponding to the current type of edge of the target node to obtain a matrix multiplication operation result;
s402: carrying out equal-scale scaling processing on the matrix multiplication result to obtain a scaling processing result;
s403: and normalizing the scaling processing result by using the normalization index function to obtain a weight coefficient corresponding to each type of edge of the target node.
As described above, in the embodiment of the present application, the weight coefficient corresponding to each type of edge of the target node is obtained through a differentiated fusion process, and the implementation manner of the differentiated fusion process includes matrix multiplication, equal scaling and normalization.
The implementation manner of the differentiated fusion process is shown in the following formula (7):
in equation (7), the information fusion is first performed on the target node (e.g. v) and the neighborhood node (denoted by o) by matrix multiplication, and h v And h o The vector expressions of the target node and the neighborhood node of the target node are respectively vectors of a hidden layer,indicating transposition. And then pass throughPerforming an equal scaling process to prevent the vector of the previous matrix multiplication from being inThe product value is too large, so that gradient disappearance or gradient explosion phenomena occur to softmax during normalization, namely, the vector inner product is scaled through scaling. The softmax normalization is to ensure that the obtained values of the weight coefficients of the same type of edges are all in the (0,1) interval after the input vector refers to the softmax function, namely S (v, o) epsilon (0,1).
When matrix multiplication operation is carried out, a second feature mapping function is introduced to serve as a linear (feature) mapping layer to carry out feature dimension conversion. On the basis of introducing the second feature mapping function, taking the map shown in fig. 4 as an example, R is obtained based on formula (7) 1(v,p) 、R 1(v,s) 、R 5(v,w) 、R 5(v,s) Weight coefficient ofRespectively as shown in the following equations (8) to (11):
wherein Q, K, L represents a second feature mapping function, e.g.Andrespectively of edge type R 1 Is referred to as a second feature mapping function, collectively as edge type R 1 A set of second feature mapping functions of (a);andrespectively of edge type R 5 Is referred to as an edge type of R 5 Is performed on the first feature mapping function.Andis for the edge type R 1 In other words, the neighborhood nodes (p and s) of the target node correspond to the node vectors at the previous time (t-1) of the target iteration number,andis for the edge type R 5 And the neighborhood nodes (w and s) of the target node correspond to the node vectors at the previous time (t-1) of the target iteration number.Is the node vector corresponding to the target node v at the previous time (t-1) of the target iteration number.
Taking equation (8) as an example, the calculation process of equation (8) is described in detail as follows:
Due to the weight coefficientIs a real number and is a real number,is a scaling process, essentially a hyper-parameter, which can be preset, for example to 8, assuming that the calculation results in0.88 shown in the following equation (12):
it should be noted that the values of the weighting coefficients are only an assumption, and in an actual operating condition, the vector calculation is a vector calculation with several hundred dimensions.
It is worth mentioning, for exampleIt represents the result of the matrix multiplication operation,indicating the result of the scaling process.
In addition, L in the second feature mapping function represents the dependency relationship between nodes and is used for representing the correlation calculation between the nodes in the high-dimensional space. For example, the correlation is typically represented in two-dimensional space using cosine distances. However, in order to satisfy the space where the coordinate systems are orthogonal in the multidimensional space, L is required for fitting, which is an N × N weight vector in relation to the target number of iterations and the correlation. The dimension of each second feature mapping function in the feature mapping function cluster can be set according to the dimension of the initial node vector and the feature dimension conversion purpose, and the numerical value of each element in the feature mapping function cluster is determined by actual data in the map structure.
Thus, the embodiment shown in fig. 6 can obtain the weight coefficient corresponding to each type of edge of the target node.
According to the object identification method provided by the embodiment of the application, the weight coefficient corresponding to each type of edge of the target node is obtained through differentiated fusion processing, such as matrix multiplication, equal-scale scaling and normalization. The method comprises the steps of introducing a set of second feature mapping functions to carry out feature dimension conversion during matrix multiplication operation, converting low dimensions into high dimensions, enabling orthogonal features to be mutually independent, selecting principal feature in subsequent calculation, not calculating full-quantity features, improving calculation accuracy and efficiency, namely, replacing an original transfer matrix mode that edge types do not share parameters by adopting a neural network shared by parameters on the edge types, enabling model parameters not to exponentially increase along with the increase of the number of nodes, reducing the number of neural networks required to be used, and reducing the calculation amount and parameter scale for constructing a behavior type prediction model. In addition, through the scaling treatment, the problems of gradient explosion and gradient disappearance are avoided.
S302: and determining feature data transmitted by different types of edges according to the weight coefficient corresponding to each type of edge and the weight parameter network model.
After the weight coefficient corresponding to each type of edge of the target node is obtained, the characteristic data transmitted by different types of edges is further obtained by combining the weight parameter network model.
For example, in obtainingThen, combining with the weight parameter network model to obtain different types of edges R 1(v,p) 、R 1(v,s) And R 5(v,w) 、R 5(v,s) Communicated characteristic dataAnd
the weight parameter network model comprises a first feature mapping function and a feature matrix, wherein the first feature mapping function is only related to the target iteration times, namely the first feature mapping function is a neural network function related to the target iteration times, is independent of the edge type and is used for feature space conversion. The feature matrix is related to both the edge type of the target node and the target iteration number, and is used for converting the information dimension during composition collection so as to align the feature dimension of the composition with the collected feature dimension.
Assume that the first mapping function is V t Setting its dimension as MxB, the feature matrix asSetting its dimension to be B × C.
In one possible design, the possible implementation manner of step S302 includes:
and obtaining a product of the weight coefficient corresponding to each type of edge, the node vector of the neighborhood node corresponding to each type of edge in the previous time of the target iteration times and the weight parameter network model corresponding to each type of edge, and further determining the obtained product as feature data transmitted by the corresponding type of edge. Wherein the product is represented by a feature vector.
Taking fig. 4 as an example, R can be calculated by the following equations (13) to (16), respectively 1(v,p) 、R 1(v,s) And R 5(v,w) 、R 5(v,s) Communicated characteristic dataAnd
for example, the aboveCalculation by substituting into equation (13)The procedure of (2) is as follows:
thereby obtainingIs a 1 × C feature vector to represent the edge type R 1(v,p) The characteristic data of the transfer.
In the same way, can obtainTo respectively represent the edge type R 1(v,s) And R 5(v,w) 、R 5(v,s) The characteristic data of the transfer.
At this point, feature data transmitted by different types of edges is determined according to the weight coefficient corresponding to each type of edge and the weight parameter network model, and the feature data is expressed by feature vectors.
Through the description of the above embodiment, the weight coefficient of each type of edge in the target iteration time is first calculated according to the graph structure, and then the feature data transmitted by the edges of different types is obtained by combining with the weight parameter network model, so that the feature data transmitted by the edges of different types of the neighborhood nodes of the target node is determined according to the graph structure, and a data basis is provided for the summary of the neighborhood information on the target node, that is, the determination of the newly added feature data of the target node in the target iteration time.
In one possible design, a possible implementation of step S202 is shown in fig. 7. Fig. 7 is a schematic flowchart of another object identification method according to an embodiment of the present application. As shown in fig. 7, the embodiment of the present application includes:
s501: acquiring the sum of the feature vectors corresponding to different types of edges of the target node to obtain the sum of the vectors;
s502: carrying out piecewise linear processing on the sum of the vectors by adopting an activation function to obtain a linear result vector;
s503: and acquiring the product of the linear mapping matrix and the linear result vector, and representing the newly added feature data of the target node in the target iteration times by using the acquired product vector.
And the feature vector corresponding to each type of edge is used for characterizing feature data transferred by each type of edge.
As described in the previous embodiments, the feature data passed by each type of edge is characterized by the feature vector corresponding to each type of edge, such as edge type R 1(v,p) The transmitted characteristic data toThe 1 × C feature vector representation. First, obtaining the sum of feature vectors corresponding to different types of edges of the target node to obtain the sum of vectors, for example, obtainingAnd summing, wherein the sum result is the sum of the vectors. And then, carrying out piecewise linear processing on the vector sum by using an activation function, wherein the activation function can be a Relu function, for example, then obtaining a product between the linear mapping matrix and a linear result vector obtained after the piecewise linear processing, and representing new feature data of the target node in the target iteration times by using a product vector representing the product.
Taking the feature data transmitted by each different type of edge of the graph shown in fig. 4 as an example, the processes of obtaining the sum of vectors, performing piecewise linear processing, and obtaining the product can be implemented by the following formula (17), so as to obtain the new feature data of the target node v shown in fig. 4 at the target iteration time (t)
Wherein, M t A linear mapping matrix is represented, which is a neural network of the fusion layer, primarily used to align vector dimensions, which may be, for example, C × M vectors. It can be understood that the dimension of the linear mapping matrix is determined by the dimension of the feature vector corresponding to each different type of edge, and the specific value of each element in the linear mapping matrix is determined by the actual data of the map structure.
The activation function such as Relu function is to increase the nonlinearity of the neural network, and by introducing the nonlinear factor, the data noise resistance of the model is increased.
When the formula (17) is operated, the linear result vector obtained by performing piecewise linear processing by using the activation function is shown as the following formula (18):
wherein for in the linear result vectorIn a word, withFor example, the operation satisfies the following formula (19) to achieve the retention of negative samples, i.e. to introduce non-linear samples.
Is an adjustable hyper-parameter, for example 0.01 may be set. As can be seen from equation (19), forWhen it is less than or equal to 0, the corresponding linear result vector does not directly take 0 to delete the negative samples, but by introducing a hyper-parameterAnd the linear result vector is kept in a certain range by the index, so that the identification accuracy of a behavior type prediction model obtained subsequently when the behavior type prediction model faces various noises is facilitated.
Substituting equation (18) and the linear mapping matrix into equation (17) yields equation (20) as shown below:
thereby representing the newly added feature data of the target node in the target iteration times by using the obtained product vector
And determining the newly added feature data of the target node during the target iteration times according to the feature data transmitted by different types of edges by the processes of obtaining the vector sum, performing piecewise linear processing and obtaining the product.
According to the object identification method provided by the embodiment of the application, after the feature data transmitted by different types of edges of the target node are obtained, the added feature data of the target node in the target iteration times is obtained by obtaining the sum of feature vectors, piecewise linear processing and product obtaining, namely, the added information is transmitted by the graph structure at the moment corresponding to the target iteration times. The piecewise linear processing adopts an activation function, so that the purpose of keeping negative samples to participate in subsequent calculation can be achieved, the data noise resistance of the behavior type prediction model can be improved, and the identification accuracy is improved.
In a possible design, possible implementations of step S203 include:
and inputting the product vector and the self hidden layer vector of the target node at the previous time of the target iteration number into a preset recurrent neural network, and determining the output as the node vector. For example byAndcalculating a node hidden layer vector at the t moment corresponding to the target iteration timesI.e. the node vector of the target node.
As described in the foregoing embodiments, for the target node, at time t, the target node hides the layerThe vector, namely the feature data to be expressed by the node vector comes from two parts, one part is newly added feature data, namely the feature dataThe other part is from the hidden layer vector coding of the target node at the previous moment (namely t-1) itself, namelyThe former is neighborhood information learned at time t, and the latter is necessary input of timing characteristics.
According to the two parts, the node vector can be obtained through a preset cyclic neural network, and the preset cyclic neural network can be a Gate controlled cyclic neural network (GRU), which is a neural network used for capturing time sequence data dependency relationship, and can store information in a long sequence and filter and predict irrelevant information. The embodiment of the present application does not explain the implementation principle of the GRU.
The implementation of the new feature data expressed by the product vector and the node vector obtained by the predetermined recurrent neural network such as GRU can be realized by the following formula (21):
wherein a predetermined recurrent neural network, e.g. GRU, receivesAndfor updatingOutputs of GRUsIs a 1 xM feature vector and an initial vector of the target nodeThe dimensions are the same.
After a target number of iterations, e.g. 5 iterations, then takeAs the final vector representation of the target node, i.e. the node vector of the target node, and then usedFor exampleThe identification of the characteristic behavior object is carried out,the represented node vectors are used to characterize the behavior type prediction model.
According to the object identification method provided by the embodiment of the application, the node vector of the target node, namely the behavior type prediction model, is obtained through the preset recurrent neural network on the basis of obtaining the newly added feature data represented by the product vector and the self hidden layer vector of the target node in the previous target iteration time. The preset cyclic neural network can store information in a long sequence and can filter information irrelevant to prediction, so that some characteristics which are long in time can be abandoned, characteristics which are more appropriate to the current time can be screened, the calculation accuracy of a behavior type prediction model obtained by the preset cyclic neural network can be improved, and accurate and efficient characteristic behavior risk recognition can be carried out on an object to be recognized through the behavior type prediction model.
Fig. 8 is a schematic structural diagram of an object recognition apparatus according to an embodiment of the present application. As shown in fig. 8, an object recognition apparatus 600 according to an embodiment of the present application includes:
the first processing module 601 is configured to acquire portrait feature data and an association relation of an object to be recognized, and generate a map structure of the object to be recognized according to the portrait feature data and the association relation, where the object to be recognized includes an enterprise object or an individual object;
the second processing module 602 is configured to obtain a characteristic behavior probability distribution of an object to be identified according to an atlas structure and a behavior type prediction model, where the behavior type prediction model is obtained by performing model training by using an edge-type parameter sharing neighborhood information fusion mechanism and the atlas structure;
the third processing module 603 is configured to determine whether the object to be identified is a characteristic behavior object according to the characteristic behavior probability distribution, where the characteristic behavior object has a characteristic behavior risk.
On the basis of fig. 8, fig. 9 is a schematic structural diagram of another object recognition apparatus provided in the embodiment of the present application. As shown in fig. 9, the object recognition apparatus 600 according to the embodiment of the present application further includes: a model learning module 604. The model learning module 604 is configured to:
determining target characteristic data of a target node according to the graph structure, wherein the target characteristic data comprises characteristic data transmitted by each neighborhood node of the target node through different types of edges, and the target node is a node representing an object to be identified in the graph structure;
determining new feature data of the target node during the target iteration times according to the feature data transmitted by different types of edges;
and obtaining a node vector according to the newly added feature data and the self-hidden layer vector of the target node in the previous target iteration time, wherein the self-hidden layer vector is used for representing the feature data of the target node self-fused in the previous target iteration time, and the node vector is used for representing the behavior type prediction model.
In one possible design, the model learning module 604 is further configured to:
determining a weight coefficient of each type of edge of the target node in the target iteration times according to the graph structure, wherein the weight coefficient is used for representing the correlation between the target node and each neighborhood node;
determining feature data transmitted by different types of edges according to the weight coefficient corresponding to each type of edge and the weight parameter network model;
the weight parameter network model comprises a first feature mapping function and a feature matrix, wherein the first feature mapping function is related to the target iteration times and used for feature space conversion, and the feature matrix is related to the edge type of the target node and the target iteration times.
In one possible design, the model learning module 604 is further configured to:
obtaining a weight coefficient corresponding to each type of edge of the target node through differentiation fusion processing according to the map structure and the feature mapping function cluster;
the sum of the weight coefficients corresponding to the edges of the same type of the target node is 1, and the feature mapping function cluster comprises a set of second feature mapping functions corresponding to the edges of each type.
In one possible design, the model learning module 604 is further configured to:
aiming at each type of edge of the target node, performing matrix multiplication operation according to the node vectors corresponding to the target node and each neighborhood node in the previous time of the target iteration times and a set of second feature mapping functions corresponding to the current type of edge of the target node to obtain a matrix multiplication operation result;
carrying out equal-scale scaling processing on the matrix multiplication operation result to obtain a scaling processing result;
and normalizing the scaling processing result by using the normalization index function to obtain a weight coefficient corresponding to each type of edge of the target node.
In one possible design, the model learning module 604 is further configured to:
and obtaining a product between the weight coefficient corresponding to each type of edge, the node vector of the neighborhood node corresponding to each type of edge in the previous time of the target iteration number and the weight parameter network model corresponding to each type of edge, and determining the obtained product as the feature data transmitted by the edge of the corresponding type.
In one possible design, the model learning module 604 is further configured to:
acquiring the sum of the feature vectors corresponding to different types of edges of a target node to obtain the sum of the vectors, wherein the feature vector corresponding to each type of edge is used for representing feature data transmitted by each type of edge;
carrying out piecewise linear processing on the sum of the vectors by adopting an activation function to obtain a linear result vector;
and acquiring the product of the linear mapping matrix and the linear result vector, and representing the newly added feature data of the target node in the target iteration times by using the acquired product vector.
In one possible design, the model learning module 604 is further configured to:
and inputting the product vector and the self hidden layer vector of the target node when the target iteration times are the previous time into a preset recurrent neural network, and determining the output as the node vector.
In one possible design, the second processing module 602 is further configured to:
mapping the node vector into a category probability distribution vector through a third feature mapping function;
and inputting the category probability distribution vector into a two-classifier to obtain a category output vector, wherein the category output vector is used for representing the characteristic behavior probability distribution of the object to be identified.
In one possible design, the third processing module 603 is further configured to:
and judging whether the object to be identified is the characteristic behavior object or not according to the category output vector and a preset category vector, wherein the preset category vector comprises a category vector of the characteristic behavior object and a category vector of a non-characteristic behavior object.
In one possible design, the first processing module 601 is further configured to:
acquiring a related object having a transaction with the object to be identified according to the portrait characteristic data of the object to be identified;
determining the relationship category between each associated object and the object to be identified according to transaction data corresponding to the transaction;
and obtaining the incidence relation of the object to be identified according to the relation category.
In one possible design, the first processing module 601 is further configured to:
and setting the object to be identified as a node of the graph structure, and setting edges of the graph structure according to the incidence relation of the object to be identified so as to generate the graph structure of the object to be identified.
In one possible design, the first processing module 601 is further configured to:
and performing feature value on the image feature data of the object to be identified, and determining the image feature data after feature value is performed as an initial vector of a target node in a graph structure.
The object identification device provided in the embodiment of the present application may perform each step of the object identification method in the above method embodiments, and the implementation principle and the technical effect are similar, which are not described herein again.
Fig. 10 is a schematic structural diagram of an electronic device according to an embodiment of the present application. As shown in fig. 10, the electronic device 700 may include: a processor 701, and a memory 702 communicatively coupled to the processor 701.
And a memory 702 for storing programs. In particular, the program may include program code comprising computer-executable instructions.
The processor 701 is configured to execute computer-executable instructions stored by the memory 702 to implement an object recognition method.
The processor 701 may be a Central Processing Unit (CPU), an Application Specific Integrated Circuit (ASIC), or one or more Integrated circuits configured to implement the embodiments of the present Application.
Alternatively, the memory 702 may be separate or integrated with the processor 701. When the memory 702 is a device separate from the processor 701, the electronic device 700 may further include:
the bus 703 is used to connect the processor 701 and the memory 702. The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended ISA (EISA) bus, or the like. Buses may be divided into address buses, data buses, control buses, etc., but do not represent only one bus or type of bus.
Alternatively, in a specific implementation, if the memory 702 and the processor 701 are implemented by being integrated on one chip, the memory 702 and the processor 701 may complete communication through an internal interface.
The present application also provides a computer-readable storage medium, which may include: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and in particular, the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used for the steps of the method in the foregoing embodiments.
The present application also provides a computer program product comprising computer executable instructions which, when executed by a processor, perform the steps of the method in the above embodiments.
Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.
It will be understood that the present application is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.
Claims (17)
1. An object recognition method, comprising:
acquiring portrait feature data and an association relation of an object to be identified, and generating a map structure of the object to be identified according to the portrait feature data and the association relation, wherein the object to be identified comprises an enterprise object or an individual object;
obtaining characteristic behavior probability distribution of the object to be recognized according to the map structure and a behavior type prediction model, wherein the behavior type prediction model is obtained by learning the map structure by adopting a side type parameter sharing neighborhood information fusion mechanism;
and determining whether the object to be identified is a characteristic behavior object according to the characteristic behavior probability distribution, wherein the characteristic behavior object has a characteristic behavior risk.
2. The object recognition method of claim 1, wherein learning the graph structure to obtain the behavior type prediction model by using a parameter sharing neighborhood information fusion mechanism of the edge type comprises:
determining target feature data of a target node according to the graph structure, wherein the target feature data comprise feature data transmitted by various neighborhood nodes of the target node through edges of different types, and the target node is a node for representing the object to be identified in the graph structure;
determining new feature data of the target node during the target iteration times according to the feature data transmitted by different types of edges;
and obtaining a node vector according to the newly added feature data and the self hidden layer vector of the target node at the previous time of the target iteration number, wherein the self hidden layer vector is used for representing the feature data self-fused to the target node at the previous time of the target iteration number, and the node vector is used for representing the behavior type prediction model.
3. The object recognition method of claim 2, wherein the determining target feature data of a target node from the graph structure comprises:
determining a weight coefficient of each type of edge of the target node in the target iteration times according to the graph structure, wherein the weight coefficient is used for representing the correlation between the target node and each neighborhood node;
determining feature data transmitted by the edges of different types according to the weight coefficient corresponding to each type of edge and a weight parameter network model;
the weight parameter network model comprises a first feature mapping function and a feature matrix, wherein the first feature mapping function is related to the target iteration times and used for feature space conversion, and the feature matrix is related to the edge type of the target node and the target iteration times.
4. The object recognition method according to claim 3, wherein the determining a weight coefficient of each type of edge of the target node at the target iteration number according to the graph structure comprises:
obtaining a weight coefficient corresponding to each type of edge of the target node through differentiation fusion processing according to the graph structure and the feature mapping function cluster;
the sum of the weight coefficients corresponding to the same type of edges of the target node is 1, and the feature mapping function cluster includes a set of second feature mapping functions corresponding to each type of edges.
5. The object recognition method according to claim 4, wherein the obtaining a weight coefficient corresponding to each type of edge of the target node through a differential fusion process according to the graph structure and the feature mapping function set comprises:
for each type of edge of the target node, performing matrix multiplication operation according to a node vector corresponding to the target node and each neighborhood node in the previous time of the target iteration times and a set of second feature mapping functions corresponding to the current type of edge of the target node to obtain a matrix multiplication operation result;
carrying out equal-scale scaling processing on the matrix multiplication result to obtain a scaling processing result;
and normalizing the scaling processing result by utilizing a normalization index function to obtain a weight coefficient corresponding to each type of edge of the target node.
6. The object recognition method according to claim 3, wherein the determining feature data delivered by the different types of edges according to the weighting coefficients corresponding to the edges of each type and a weighting parameter network model comprises:
and obtaining a product between the weight coefficient corresponding to each type of edge, the node vector of the neighborhood node corresponding to each type of edge in the previous time of the target iteration times and the weight parameter network model corresponding to each type of edge, and determining the obtained product as feature data transmitted by the edge of the corresponding type.
7. The object recognition method according to any one of claims 2 to 6, wherein the determining new feature data of the target node at the target iteration number according to the feature data transferred by each different type of edge comprises:
acquiring the sum of the feature vectors corresponding to different types of edges of the target node to obtain the sum of the vectors, wherein the feature vector corresponding to each type of edge is used for representing feature data transmitted by each type of edge;
performing piecewise linear processing on the vector sum by adopting an activation function to obtain a linear result vector;
and acquiring the product of the linear mapping matrix and the linear result vector, and representing the newly added feature data of the target node when the target iteration times is carried out by using the acquired product vector.
8. The object recognition method according to claim 7, wherein the obtaining a node vector according to the newly added feature data and a self-hidden layer vector of the target node at a previous time of the target iteration number comprises:
and inputting the product vector and the self hidden layer vector of the target node when the target iteration times is the previous time into a preset recurrent neural network, and determining the output as the node vector.
9. The object recognition method according to any one of claims 2 to 6, wherein obtaining the probability distribution of the object to be recognized according to the graph structure and behavior type prediction model comprises:
mapping the node vector into a category probability distribution vector through a third feature mapping function;
and inputting the category probability distribution vector into a classifier to obtain a category output vector, wherein the category output vector is used for representing the characteristic behavior probability distribution of the object to be identified.
10. The object recognition method according to claim 9, wherein the determining whether the object to be recognized is a feature behavior object according to the feature behavior probability distribution includes:
and judging whether the object to be identified is the characteristic behavior object or not according to the category output vector and a preset category vector, wherein the preset category vector comprises a category vector of the characteristic behavior object and a category vector of a non-characteristic behavior object.
11. The object recognition method according to any one of claims 2 to 6, wherein obtaining the association relationship of the object to be recognized comprises:
acquiring an associated object having a transaction with the object to be identified according to the portrait characteristic data of the object to be identified;
determining the relationship category between each associated object and the object to be identified according to the transaction data corresponding to the transaction traffic;
and obtaining the incidence relation of the object to be identified according to the relation category.
12. The object recognition method of claim 11, wherein the generating of the atlas structure of the object to be recognized according to the portrait feature data and the association relation comprises:
setting the object to be identified as a node of the graph structure, and setting the edge of the graph structure according to the incidence relation of the object to be identified so as to generate the graph structure of the object to be identified.
13. The object recognition method according to any one of claims 2 to 6, further comprising, after the acquiring of the portrait feature data of the object to be recognized:
and performing feature value on the image feature data of the object to be recognized, and determining the image feature data after feature value is performed as an initial vector of the target node in the graph structure.
14. An object recognition apparatus, comprising:
the identification system comprises a first processing module, a second processing module and a recognition module, wherein the first processing module is used for acquiring portrait characteristic data and an incidence relation of an object to be identified, and generating a map structure of the object to be identified according to the portrait characteristic data and the incidence relation, and the object to be identified comprises an enterprise object or an individual object;
the second processing module is used for obtaining the characteristic behavior probability distribution of the object to be identified according to the map structure and the behavior type prediction model, wherein the behavior type prediction model is obtained by learning the map structure by adopting a side type parameter sharing neighborhood information fusion mechanism;
and the third processing module is used for determining whether the object to be identified is a characteristic behavior object according to the characteristic behavior probability distribution, wherein the characteristic behavior object has a characteristic behavior risk.
15. An electronic device, comprising: a processor, and a memory communicatively coupled to the processor;
the memory stores computer execution instructions;
the processor executes computer-executable instructions stored by the memory to implement the object recognition method of any one of claims 1 to 13.
16. A computer-readable storage medium having computer-executable instructions stored thereon, which when executed by a processor, are configured to implement the object recognition method of any one of claims 1 to 13.
17. A computer program product comprising computer executable instructions for implementing an object recognition method as claimed in any one of claims 1 to 13 when executed by a processor.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211261970.8A CN115511606A (en) | 2022-10-14 | 2022-10-14 | Object identification method, device, equipment and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211261970.8A CN115511606A (en) | 2022-10-14 | 2022-10-14 | Object identification method, device, equipment and storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115511606A true CN115511606A (en) | 2022-12-23 |
Family
ID=84510826
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211261970.8A Pending CN115511606A (en) | 2022-10-14 | 2022-10-14 | Object identification method, device, equipment and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115511606A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117056594A (en) * | 2023-07-31 | 2023-11-14 | 中移互联网有限公司 | User identification method and device based on interaction relationship and electronic equipment |
-
2022
- 2022-10-14 CN CN202211261970.8A patent/CN115511606A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117056594A (en) * | 2023-07-31 | 2023-11-14 | 中移互联网有限公司 | User identification method and device based on interaction relationship and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110188198B (en) | Anti-fraud method and device based on knowledge graph | |
CN111860573A (en) | Model training method, image class detection method and device and electronic equipment | |
US20210303970A1 (en) | Processing data using multiple neural networks | |
CN112347367B (en) | Information service providing method, apparatus, electronic device and storage medium | |
US11227217B1 (en) | Entity transaction attribute determination method and apparatus | |
CN111061856A (en) | Knowledge perception-based news recommendation method | |
CN111428217B (en) | Fraudulent party identification method, apparatus, electronic device and computer readable storage medium | |
CN110852881A (en) | Risk account identification method and device, electronic equipment and medium | |
CN111371767A (en) | Malicious account identification method, malicious account identification device, medium and electronic device | |
US20230049817A1 (en) | Performance-adaptive sampling strategy towards fast and accurate graph neural networks | |
CN114255121A (en) | Credit risk prediction model training method and credit risk prediction method | |
CN115511606A (en) | Object identification method, device, equipment and storage medium | |
Basak et al. | Causal ordering and inference on acyclic networks | |
CN113378090B (en) | Internet website similarity analysis method and device and readable storage medium | |
CN113408582B (en) | Training method and device for feature evaluation model | |
Khan et al. | Enhancement of neural networks model’s predictions of currencies exchange rates by phase space reconstruction and Harris Hawks’ optimization | |
CN115730125A (en) | Object identification method and device, computer equipment and storage medium | |
CN111507461A (en) | Interpretability information determining method and device | |
CN113222609B (en) | Risk identification method and device | |
US20230088840A1 (en) | Dynamic assessment of cryptocurrency transactions and technology adaptation metrics | |
CN111291838B (en) | Method and device for interpreting entity object classification result | |
CN113361652A (en) | Individual income prediction oriented depolarization method and device | |
CN112597390A (en) | Block chain big data processing method based on digital finance and big data server | |
US20240161117A1 (en) | Trigger-Based Electronic Fund Transfers | |
CN112967134B (en) | Network training method, risk user identification method, device, equipment and 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 |