Object recognition method, apparatus, device, and storage medium

Abstract

The application provides an object recognition method, device and equipment and a storage medium. The portrait feature data and the correlation of the object to be recognized are acquired to generate a graph structure of the object to be recognized, and then a behavior type prediction model learned by a parameter sharing neighborhood information fusion mechanism of an edge type is used to determine the feature behavior probability distribution of the object to be recognized. Whether the object to be recognized is a feature behavior object is determined according to the feature behavior probability distribution. The behavior type prediction model obtained by the parameter sharing neighborhood information fusion mechanism of the edge type is suitable for a graph with a large number of nodes, reduces the number of neural networks required for model construction, reduces the model calculation amount and size, avoids the gradient vanishing problem and the gradient explosion problem, 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 a feature behavior risk is accurately and efficiently recognized.

Description

Technical Field

[0001] This application relates to the field of financial technology (Fintech), and more particularly to an object recognition method, apparatus, device, and storage medium. Background Technology

[0002] With the rapid development of computer and internet technologies, Fintech, as a product of the deep integration of finance and technology, is currently becoming a hot topic for innovation and development in the financial industry. For financial institutions, identifying whether a business entity exhibits characteristic behaviors that pose risks, such as fraud, is crucial for their own security and other aspects.

[0003] Identifying whether an enterprise possesses characteristic behavioral risks is typically achieved using deep learning models based on knowledge graphs. However, existing technologies, when constructing learning models to compute neighborhood information propagation, have a relationship between the number of neural networks to be trained and the size of the knowledge graph (number of nodes): Number of neural networks = O(|number of nodes| 2 As can be seen, existing technologies require too many neural networks to train, and the memory and computational resources used in training increase exponentially with the number of nodes. This leads to increased computational load and decreased training efficiency, making them unsuitable for graphs with a large number of nodes. Furthermore, excessive neural networks are prone to gradient vanishing or exploding problems during T iterations of training due to the large number of nodes, resulting in model training failure and reduced training and prediction efficiency.

[0004] Furthermore, in using existing deep learning models based on knowledge graphs, it was found that the subgraphs partitioned by different types of edges exhibit significant variance in both the number of edges and nodes. This variance is caused by the uneven distribution of edges across nodes. The existing method of sorting nodes and using a shared transition matrix for model training and iterative computation results in many neural networks not being trained, leading to poor predictive ability and inaccurate predictions. Summary of the Invention

[0005] This application provides an object recognition method, apparatus, device, and storage medium to solve the technical problems of low model training efficiency, low prediction efficiency, and poor prediction ability when using deep learning models built based on graphs for object recognition in the prior art.

[0006] Firstly, this application provides an object recognition method, including:

[0007] Acquire the portrait feature data and association relationships of the object to be identified, and generate the map structure of the object to be identified based on the portrait feature data and the association relationships. The object to be identified includes enterprise objects or individual objects.

[0008] Based on the graph structure and behavior type prediction model, the feature behavior probability distribution of the object to be identified is obtained. The behavior type prediction model is learned by learning the graph structure using a parameter-sharing neighborhood information fusion mechanism of edge type.

[0009] Based on the probability distribution of the characteristic behaviors, it is determined whether the object to be identified is a characteristic behavior object, and the characteristic behavior object has characteristic behavior risk.

[0010] In one possible design, a parameter-sharing neighborhood information fusion mechanism for the edge type is used to learn the graph structure to obtain the behavior type prediction model, including:

[0011] The target feature data of the target node is determined based on the graph structure. The target feature data includes feature data transmitted by each neighboring node of the target node through different types of edges. The target node is the node in the graph structure that represents the object to be identified.

[0012] Based on the feature data transmitted by different types of edges, determine the new feature data of the target node at the target iteration number;

[0013] The node vector is obtained based on the newly added feature data and the target node's own hidden layer vector at the previous target iteration number. The self-hidden layer vector is used to represent the feature data that the target node self-fused to at the previous target iteration number, and the node vector is used to represent the behavior type prediction model.

[0014] In one possible design, determining the target feature data of the target node based on the graph structure includes:

[0015] The weight coefficients of each type of edge of the target node are determined according to the graph structure at the target iteration number, and the weight coefficients are used to characterize the correlation between the target node and each of the neighboring nodes;

[0016] The feature data transmitted by each type of edge is determined based on the weight coefficients and weight parameter network model corresponding to each type of edge.

[0017] The weighted parameter network model includes a first feature mapping function and a feature matrix. The first feature mapping function is related to the target iteration number and is used for feature space transformation. The feature matrix is ​​related to the edge type of the target node and the target iteration number.

[0018] In one possible design, determining the weight coefficients of each type of edge of the target node at the target iteration number based on the graph structure includes:

[0019] Based on the graph structure and the feature mapping function cluster, the weight coefficients corresponding to each type of edge of the target node are obtained through differential fusion processing;

[0020] Wherein, 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 edge;

[0021] In one possible design, obtaining the weight coefficients for each type of edge of the target node through differential fusion processing based on the graph structure and the set of feature mapping functions includes:

[0022] For each type of edge of the target node, matrix multiplication is performed based on the node vectors corresponding to the target node and each of the neighboring nodes in the previous iteration of the target iteration, and the set of second feature mapping functions corresponding to the edge of the current type of the target node, to obtain the matrix multiplication result.

[0023] The matrix multiplication result is scaled proportionally to obtain the scaled result.

[0024] The scaling result is normalized using a normalized exponential function to obtain the weight coefficients corresponding to each type of edge of the target node.

[0025] In one possible design, determining the feature data transmitted by the different types of edges based on the weight coefficients and weight parameter network model corresponding to each type of edge includes:

[0026] The product of the weight coefficients corresponding to each type of edge, the node vectors of the neighboring nodes corresponding to each type of edge before the target iteration number, and the weight parameters of the network model corresponding to each type of edge is obtained, and the resulting product is determined as the feature data transmitted by the edge of the corresponding type.

[0027] In one possible design, determining the new feature data of the target node at the target iteration number based on the feature data passed by each type of edge includes:

[0028] The sum of feature vectors corresponding to each type of edge of the target node is obtained, and the feature vector corresponding to each type of edge is used to characterize the feature data transmitted by each type of edge.

[0029] The sum of the vectors is processed piecewise linearly using an activation function to obtain a linear result vector;

[0030] Obtain the product of the linear mapping matrix and the linear result vector, and use the resulting product vector to represent the newly added feature data of the target node at the target iteration number.

[0031] In one possible design, obtaining the node vector based on the newly added feature data and the target node's own hidden layer vector at the previous target iteration number includes:

[0032] The product vector and the hidden layer vector of the target node at the previous iteration number are input into a preset recurrent neural network, and the output is determined as the node vector.

[0033] In one possible design, the probability distribution of the object to be identified is obtained based on the graph structure and behavior type prediction model, including:

[0034] The node vector is mapped to a category probability distribution vector using a third feature mapping function;

[0035] The category probability distribution vector is input into a binary classifier to obtain a category output vector, which is used to characterize the feature behavior probability distribution of the object to be identified.

[0036] In one possible design, determining whether the object to be identified is a characteristic behavior object based on the characteristic behavior probability distribution includes:

[0037] The object to be identified is determined based on the category output vector and the preset category vector. The preset category vector includes the category vector of the characteristic behavior object and the category vector of non-characteristic behavior objects.

[0038] In one possible design, obtaining the association relationship of the object to be identified includes:

[0039] Based on the profile feature data of the object to be identified, obtain the associated objects that have transactional relationships with the object to be identified;

[0040] The relationship category between each associated object and the object to be identified is determined based on the transaction data corresponding to the transactions.

[0041] The association relationship of the object to be identified is obtained based on the relationship category.

[0042] In one possible design, generating the atlas structure of the object to be identified based on the portrait feature data and the association relationship includes:

[0043] The object to be identified is set as a node of the graph structure, and the edges of the graph structure are set according to the association relationship of the object to be identified, so as to generate the graph structure of the object to be identified.

[0044] In one possible design, after acquiring the portrait feature data of the object to be identified, the following is also included:

[0045] The eigenvalued portrait feature data of the object to be identified is used as the initial vector of the target node in the graph structure.

[0046] Secondly, this application provides an object recognition device, comprising:

[0047] The first processing module is used to acquire the portrait feature data and association relationships of the object to be identified, and generate the map structure of the object to be identified based on the portrait feature data and the association relationships. The object to be identified includes enterprise objects or individual objects.

[0048] The second processing module is used to obtain the feature behavior probability distribution of the object to be identified based on the graph structure and behavior type prediction model. The behavior type prediction model is obtained by learning the graph structure using a parameter-sharing neighborhood information fusion mechanism of edge type.

[0049] The third processing module is used to determine whether the object to be identified is a characteristic behavior object based on the characteristic behavior probability distribution, and the characteristic behavior object has characteristic behavior risk.

[0050] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0051] The memory stores computer-executed instructions;

[0052] The processor executes computer execution instructions stored in the memory to implement any of the possible object recognition methods provided in the first aspect.

[0053] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement any of the possible object recognition methods provided in the first aspect.

[0054] Fifthly, this application provides a computer program product, including computer execution instructions, which, when executed by a processor, are used to implement any of the possible object recognition methods provided in the first aspect.

[0055] This application provides an object recognition method, apparatus, device, and storage medium. First, it acquires the portrait feature data and association relationships of the object to be identified. Then, it generates a graph structure of the object to be identified based on the portrait feature data and association relationships. Next, based on the graph structure and a behavior type prediction model, it obtains the probability distribution of the characteristic behaviors of the object to be identified. The behavior type prediction model is trained using an edge-type parameter-sharing neighborhood information fusion mechanism and the graph structure. Finally, it determines whether the object to be identified is an object exhibiting characteristic behaviors based on the probability distribution of characteristic behaviors. The behavior type prediction model is constructed using an edge-type parameter-sharing neighborhood information fusion mechanism. The model parameters do not increase exponentially with the number of nodes, thus it is applicable to graphs with a large number of nodes. It also reduces the number of neural networks required for model learning, effectively reducing the computational load and scale of the model, avoiding the vanishing and exploding gradient problems, and improving learning and prediction efficiency. Furthermore, the edge-type parameter-sharing neighborhood information fusion mechanism increases the distribution differences between neighboring nodes through differentiated fusion, allowing the constructed behavior type prediction model to be fully learned and accurately and efficiently identify objects with characteristic behavioral risks. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of this application;

[0058] Figure 2 A flowchart illustrating an object recognition method provided in an embodiment of this application;

[0059] Figure 3 A flowchart illustrating another object recognition method provided in an embodiment of this application;

[0060] Figure 4 This is a partial schematic diagram of a spectrum structure provided in an embodiment of this application;

[0061] Figure 5 A flowchart illustrating another object recognition method provided in an embodiment of this application;

[0062] Figure 6 A flowchart illustrating yet another object recognition method provided in an embodiment of this application;

[0063] Figure 7A flowchart illustrating yet another object recognition method provided in an embodiment of this application;

[0064] Figure 8 This is a schematic diagram of the structure of an object recognition device provided in an embodiment of this application;

[0065] Figure 9 This is a schematic diagram of another object recognition device provided in an embodiment of this application;

[0066] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0067] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and apparatus consistent with some aspects of this application as detailed in the appended claims.

[0068] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0069] Identifying whether a business or individual entity possesses characteristic behavioral risks is typically achieved using deep learning models based on knowledge graphs. However, existing technologies, when constructing learning models to compute neighborhood information propagation, have a relationship between the number of neural networks to be trained and the size of the knowledge graph (number of nodes): Number of neural networks = O(|number of nodes| 2It is evident that existing technologies require training too many neural networks, and the memory and computational resources used during training increase exponentially with the number of nodes, leading to increased computational load and decreased training efficiency, making them unsuitable for graphs with a large number of nodes. Furthermore, excessive neural networks, due to the large number of nodes during T iterations of training, are prone to gradient vanishing or exploding problems, resulting in model training failure and reduced training and prediction efficiency. In addition, during the use of existing deep learning models based on knowledge graphs, it was found that the subgraphs partitioned by different types of edges exhibit significant variance in both the number of edges and nodes. This variance is caused by the uneven distribution of edges across nodes. Existing methods that sort nodes and use a shared transition matrix for model training and iterative computation result in many neural networks not being trained, leading to poor predictive ability and inaccurate predictions.

[0070] To address the aforementioned problems in existing technologies, this application provides an object recognition method, apparatus, device, and storage medium. The inventive concept of the object recognition method provided in this application lies in: proposing a parameter-sharing neighborhood information fusion mechanism based on edge types to construct a behavior type prediction model. When calculating the propagation of neighborhood information, different types of edges are distinguished, and the information propagation, i.e., the transmitted feature data, on each neighboring node of the target node is calculated according to the edge type. This allows the use of a neural network with parameter sharing on edge types to replace the original method of using transfer matrices for different parameters of edge types. This prevents the model parameters from increasing exponentially with the number of nodes, making the behavior type prediction model construction method suitable for graphs with a large number of nodes. Furthermore, it reduces the number of neural networks required for model training, effectively reducing the computational load and training scale, avoiding gradient vanishing and gradient exploding problems, and improving training and prediction efficiency. In addition, the parameter-sharing neighborhood information fusion mechanism based on edge types can reflect the uneven distribution of edges on nodes in the graph through the weight coefficients of different edge types. This fully considers the distribution differences between neighboring nodes, ensuring that the trained behavior type prediction model is sufficiently trained, thereby effectively improving the accuracy of object recognition.

[0071] The following describes exemplary application scenarios of the embodiments of this application.

[0072] Figure 1 This is a schematic diagram of an application scenario provided in an embodiment of this application, such as... Figure 1As shown, the network serves as a medium for providing a communication link between the first electronic device 100 and the second electronic device 200. The network can include various connection types, such as wired, wireless communication links, or fiber optic cables. The first electronic device 100 and the second electronic device 200 can interact through the network to receive or send messages. The first electronic device 100 can be any terminal capable of obtaining profile feature data and related relationships of the object to be identified. The first electronic device 100 can be configured at the object to be identified itself or at a third-party institution, etc., and this embodiment does not limit this. The second electronic device 200 can be deployed at financial institutions, etc., and is configured to execute the object identification method provided in this application embodiment. The second electronic device 200 obtains the profile feature data and related relationships of the object to be identified from the first electronic device 100, and then executes the object identification method provided in this application embodiment to identify whether the object to be identified is a characteristic behavior object.

[0073] Optionally, the objects to be identified may include various enterprises, organizations, or individuals, while the objects with characteristic behaviors refer to enterprises, organizations, or individuals with characteristic behavioral risks, such as fraud or deception risks.

[0074] It should be noted that the embodiments of this application do not limit the type of the first electronic device 100 described above. For example, the first electronic device 100 may be a computer, smartphone, smart glasses, smart bracelet, smartwatch, tablet computer, etc. Figure 1 The first electronic device 100 is illustrated as a computer. The second electronic device 200 can be a server, a computer, a server cluster, etc. Figure 1 The example shown is a Chinese server.

[0075] It should be noted that the above application scenarios are merely illustrative, and the object recognition method provided in this application includes, but is not limited to, the above application scenarios.

[0076] Figure 2 This is a flowchart illustrating an object recognition method provided in an embodiment of this application. Figure 2 As shown, the object recognition method provided in this application includes:

[0077] S101: Obtain the portrait feature data and correlation of the object to be identified, and generate the map structure of the object to be identified based on the portrait feature data and correlation.

[0078] Data is acquired to characterize the independent attributes of the object to be identified, such as the object's profile features, as well as data to characterize the hidden relationships between the object and other objects, such as the association relationships between the objects. Then, a graph structure of the object to be identified is constructed based on the object's profile features and association relationships.

[0079] Taking an enterprise as the target of identification as an example, the profile feature data of the target of identification may include, for example, basic enterprise information such as industry category, legal person information, registered address, legal person type, and technical number; enterprise registration period, number of changes in business registration information, number of changes in legal person information, and other enterprise business category information; enterprise loan / deposit balance, monthly deposit accumulation, and other enterprise deposit information; transfer information such as number of transfers out within half a year, amount of transfers out within half a year, number of transfers in within half a year, amount of transfers out within half a year, and transfers out within half a year; as well as credit information such as enterprise blacklists and enterprise credit reports. The specific content of the profile features of the target of identification is determined by the type of the target of identification, and this application embodiment does not limit this.

[0080] In one possible design, the possible implementation methods for obtaining the association relationship of the object to be identified in step S101 include:

[0081] Based on the profile feature data of the object to be identified, related objects with transaction relationships with the object to be identified are obtained. Then, the relationship category between each related object and the object to be identified is determined according to the transaction data corresponding to the transaction. The association relationship of the object to be identified is then determined according to the relationship category. One relationship category corresponds to one association relationship.

[0082] For example, if the object to be identified is a company, the determined relationship categories may include, but are not limited to, the following:

[0083] The first type involves upstream and downstream transfers between companies; the second type involves shared key stakeholders between companies, such as legal person relationships, guarantor relationships, stakeholder relationships, and close kinship relationships between stakeholders; the third type involves trade relationships based on frequency, amount, single transaction amount, number of transactions, transaction time range, and transaction location; the fourth type involves equity relationships where companies hold controlling stakes in each other; the fifth type involves shared profile data such as registered address, phone number, emergency contact, bank account, and company email; and the sixth type involves customer relationships with similar customer groups.

[0084] In one possible design, step S101 generates the atlas structure of the object to be identified based on the portrait feature data and correlations, including:

[0085] The objects to be identified are set as nodes in the graph structure, and the edges of the graph structure are set according to the relationships between the objects. That is, one edge represents one relationship type, thus constructing the graph structure of the objects to be identified. After obtaining the portrait feature data of the objects to be identified, the portrait feature data is eigenvalued, and then the eigenvalued portrait feature data is used as the initial vector of the target nodes in the graph structure, which can be represented as... This refers to the feature expression of the target node before it has started the target iteration count. v represents the target node, which is the node in the graph structure representing the object to be identified.

[0086] It should be noted that the initial vector is 1×M dimensional. To achieve better classification results during training, a higher-dimensional feature space can be used. For example, the dimension of the target node can be set to 300 dimensions. The value in each dimension represents a portrait feature data point after the feature values ​​are calculated. If the number of portrait feature data points is less than the dimension of the initial vector, the value in that dimension can be placed as 0.

[0087] S102: Based on the graph structure and behavior type prediction model, obtain the characteristic behavior probability distribution of the object to be identified.

[0088] Among them, the behavior type prediction model is obtained by learning the graph structure using a parameter-sharing neighborhood information fusion mechanism based on edge types.

[0089] After constructing the graph structure of the object to be identified, a behavior type prediction model is obtained by using this graph structure and the parameter-sharing neighborhood information fusion mechanism of edge types for training. In other words, the behavior type prediction model is constructed by training the graph results using the parameter-sharing neighborhood information fusion mechanism of edge types. Therefore, the behavior type prediction model is a feature representation of the object to be identified. Then, the probability distribution of the feature behaviors of the object to be identified is obtained through the behavior type prediction model and a binary classifier, thus enabling the prediction of whether the object to be identified is a feature behavior object.

[0090] In the process of learning the behavior type prediction model from the graph structure using a parameter-sharing neighborhood information fusion mechanism based on edge types, the feature data transmitted by different types of edges can be considered, allowing for differentiated propagation of neighborhood information of the target node. Furthermore, by obtaining feature data transmitted by each neighboring node based on different edge types, different edge types can distinguish different relationships between the objects to be identified. These different relationships manifest as different feature extraction networks in the model construction. This is equivalent to using a neural network with parameter sharing across edge types to replace the existing method of using a transition matrix with non-shared parameters across edge types. This prevents the model parameters from increasing exponentially with the number of nodes during model construction. Parameter sharing across edge types also reduces the number of neural networks required for model construction, thereby reducing the computational load and training scale, resulting in faster model convergence. It also avoids the vanishing and exploding gradient problems, improving both training and prediction efficiency.

[0091] In one possible design, step S102 employs a parameter-sharing neighborhood information fusion mechanism based on edge types to learn the graph structure and obtain the behavior type prediction model. Possible implementation methods include... Figure 3 As shown. Figure 3 This is a flowchart illustrating another object recognition method provided in an embodiment of this application. Figure 3 As shown, the embodiments of this application include:

[0092] S201: Determine the target feature data of the target node based on the graph structure.

[0093] The target feature data includes feature data transmitted by each neighboring node of the target node through different types of edges. The target node is the node in the graph structure that represents the object to be identified.

[0094] Based on the graph structure, the feature data of each neighboring node of the target node in the graph structure is obtained through different types of edges, that is, the target feature data of the target node is determined according to the graph structure.

[0095] For example, the neighborhood structure of the target node v in the graph structure is as follows: Figure 4 As shown, the target node v has three neighboring nodes p, w, and s, and these nodes are associated with each other: R 1(v,p) R 1(v,s) and R 5(v,w) R 5(v,s) Among them, R 1(v,p) This indicates that the target node v, representing the object to be identified, and the related object p (represented by neighboring nodes) that has a transaction relationship with the object to be identified have a relationship of type R1, and so on, R... 1(v,s) and R 5(v,w) R5(v,s) .

[0096] Therefore, this step calculates the feature data propagated by each neighboring node {p,s,w} of the target node through different edge types, such as {R1,R5}, based on the graph results. and That is, the target feature data of the target node. and In the process of building the model, there are a fixed number of iterations, and each iteration is called information propagation at time t, or feature data transfer. After multiple iterations of calculation, the feature expression of the target node is regarded as the node vector after the target node has learned the graph structure.

[0097] S202: Determine the new feature data of the target node at the target iteration number based on the feature data passed by each type of edge.

[0098] After obtaining the target feature data of the target node, that is, the feature data of different types of edge propagation of the target node, the newly added feature data of the target node at the target iteration number is determined based on the feature data of different types of edge propagation. That is, the information added by the graph structure through propagation at time t, for example, using This means that at time t, the target node v is surrounded by edges of different types (e.g., R). 1(v,p) R 1(v,s) and R 5(v,w) R 5(v,s) This involves summarizing newly added neighborhood information on a graph composed of isomorphic graphs. In other words, this step is achieved through... and calculate

[0099] S203: Obtain the node vector based on the newly added feature data and the target node's own hidden layer vector at the previous iteration number.

[0100] Among them, the self-hidden layer vector is used to represent the feature data that the target node self-fused to in the previous target iteration, and the node vector is used to represent the behavior type prediction model.

[0101] For the target node, at time t, the target node's hidden layer The feature data to be expressed by the vector, that is, the node vector, comes from two parts: one part is newly added feature data, that is... The other part comes from the hidden layer vector encoding of the target node itself at the previous time step (i.e., t-1). Therefore, this step involves adding new feature data. and its own hidden vector Calculate the hidden layer vector of the target node at time t. In other words, node vectors. The calculated node vectors represent the behavior type prediction model.

[0102] The object recognition method provided in this application first obtains the feature data transmitted by each neighboring node through different types of edges, and then obtains the new feature data of the target node accordingly. Then, it combines the target node's own hidden vector at the previous iteration number to obtain the node vector used to express the behavior type prediction model, completes the learning of the graph structure, and obtains the behavior type prediction model that predicts whether the object to be identified is a feature behavior object.

[0103] Based on the obtained behavior type prediction model, the possible implementation methods for obtaining the probability distribution of the object to be identified in step S102 according to the graph structure and behavior type prediction model include:

[0104] First, the node vector is mapped to a category probability distribution vector through the third feature mapping function. Then, the category probability distribution vector is input into the binary classifier to obtain the category output vector, which is used to characterize the feature behavior probability distribution of the object to be identified.

[0105] To predict whether an object to be identified is a characteristic behavior object, a classifier can be used to classify the feature representation of the object to be identified. For example, the characteristic behavior probability distribution can be obtained using the following formula (1):

[0106]

[0107] in, The output vector represents the category; the sigmoid function represents a binary classifier. This represents the category probability distribution vector.

[0108] Specifically, since the Sigmoid function is highly efficient with Bernoulli distributions on probabilities q and (1-q), and its output range is (0,1), the binary classifier uses the Sigmoid function. And in order to make the node vector... If the data can be input into a binary classifier, then the dimensions of the node vectors need to be mapped using features. For example, a third feature mapping function represented by S′ can be used to map the 1×M dimensions. The vector is mapped to a 1×2 class probability distribution vector.

[0109] For example, the mapping process using the third feature mapping function is shown in the following formula (2):

[0110]

[0111] in, It is an M-dimensional feature representation f; the third feature mapping function S′ is an M×2 dimensional vector; This represents the category probability distribution vector.

[0112] The Sigmoid function is calculated as shown in formula (3):

[0113]

[0114] Based on the above formulas (1) to (3), we can obtain the following formula (4):

[0115]

[0116] The meaning expressed by the feature behavior probability distribution represented by the category output vector is shown in the following formula (5):

[0117]

[0118] It should be noted that the binary classifier used to obtain the category output vector is the trained binary classifier. That is, before inputting the category probability distribution vector into the binary classifier, the binary classifier needs to be trained with training samples so that the trained binary classifier can identify the object to be identified as a feature behavior object or a non-feature behavior object.

[0119] Alternatively, the binary classifier can be trained using the loss function shown in formula (6) below:

[0120]

[0121] Where L represents the loss function, D l The term represents a sample object, which includes sample objects with characteristic behaviors and sample objects without characteristic behaviors. y′ represents the label of the object represented by node u, which is a sample object with characteristic behaviors. The binary classifier predicts that node u represents the label of the feature behavior object.

[0122] As can be seen from formula (6), the loss value of the binary classifier is obtained by accumulating the maximum likelihood estimates of all nodes with characteristic behaviors. This loss value is used to train the binary classifier, and the parameters are adjusted during training to fit the data features. The loss function is the cross-entropy function.

[0123] S103: Determine whether the object to be identified is a characteristic behavior object based on the probability distribution of characteristic behaviors.

[0124] After obtaining the behavior probability distribution of the object to be identified, it is determined whether the object to be identified is a characteristic behavior object based on the behavior probability distribution of the object to be identified. Characteristic behavior objects have characteristic behavior risks, thereby identifying whether the object to be identified has characteristic behavior risks.

[0125] For example, the category output vector representing the probability distribution of characteristic behaviors is compared with a preset category vector to determine whether the object to be identified is a characteristic behavior object. The preset category vector includes category vectors for characteristic behavior objects and category vectors for non-characteristic behavior objects.

[0126] For example, the category vector for objects exhibiting characteristic behavior might be [1,0], while the category vector for objects exhibiting non-characteristic behavior might be [0,1]. If the category output vector is the former, it indicates that the object to be identified is a characteristic behavior object, meaning that the object to be identified has a characteristic behavior risk. If the category output vector is the latter, it indicates that the object to be identified is a non-characteristic behavior object, meaning that the object to be identified does not have a characteristic behavior risk.

[0127] Understandably, the preset category vectors are obtained through semantic annotation during the training of the binary classifier, and their purpose is to interpret the semantics of the binary classifier's output. For example, the annotation method is as follows:

[0128] If the object is a feature behavior object, the category is 1, and the corresponding category vector is [1,0]; if it is not a feature behavior object, the category vector is [0,1]. This application does not limit the annotation method; it can be set according to the specific recognition task.

[0129] Thus, the object recognition method provided in this application embodiment can efficiently and accurately identify whether an object to be identified is a characteristic behavior object. Table 1 shows the accuracy of constructing behavior type prediction models using existing technologies and the object recognition method provided in this application embodiment, as well as the accuracy of identifying the object to be identified.

[0130] Table 1

[0131]

[0132] As shown in Table 1, the model construction accuracy and recognition accuracy of the object recognition method provided in this application are both higher than those of the prior art. The difference in recognition accuracy is significant, indicating that the behavior type prediction model constructed in the object recognition method provided in this application has stronger generalization ability. Specifically, this means that the behavior type prediction model has been trained and learned more thoroughly, and the behavior type prediction model shares parameters across edge types. Sufficient training and learning can effectively adjust model parameters, which helps improve prediction ability. Furthermore, the differentiated neighborhood fusion method can widen the distinction between different neighborhood information, thereby increasing the difference between neighborhood information and improving the generalization ability of the behavior type prediction model.

[0133] Table 2 shows the computational and memory resources consumed when constructing a behavior type prediction model for iterative calculation using the object recognition method provided by the prior art and the embodiments of this application.

[0134] Table 2

[0135]

[0136] As can be seen from Table 2, the existing solutions consume a lot of CPU resources due to the presence of a large number of redundant parameters, and the excessive number of parameters requires a large amount of memory resources to store variables and intermediate values.

[0137] The object recognition method provided in this application first acquires the portrait feature data and association relationships of the object to be identified, then generates a graph structure of the object to be identified based on the portrait feature data and association relationships, and then obtains the characteristic behavior probability distribution of the object to be identified based on the graph structure and behavior type prediction model. The behavior type prediction model is trained using an edge-type parameter-sharing neighborhood information fusion mechanism and the graph structure. Finally, it determines whether the object to be identified is a characteristic behavior object based on the characteristic behavior probability distribution. The behavior type prediction model is constructed using an edge-type parameter-sharing neighborhood information fusion mechanism. The model parameters do not increase exponentially with the number of nodes, thus it is applicable to graphs with a large number of nodes and reduces the number of neural networks required for model training, thereby effectively reducing the computational load and training scale, avoiding gradient vanishing and gradient exploding problems, and improving training and prediction efficiency. Furthermore, the edge-type parameter-sharing neighborhood information fusion mechanism increases the distribution differences between neighboring nodes through differential fusion, allowing the constructed behavior type prediction model to be fully trained and accurately and efficiently identify objects with characteristic behavior risks.

[0138] In one possible design, step S201 could be implemented as follows: Figure 5 As shown. Figure 5 This is a flowchart illustrating another object recognition method provided in an embodiment of this application. Figure 5 As shown, the embodiments of this application include:

[0139] S301: Determine the weight coefficient of each type of edge of the target node at the target iteration number based on the graph structure.

[0140] The weighting coefficient is used to characterize the correlation between the target node and its neighboring nodes.

[0141] Based on the graph structure, determine the weight coefficient of each type of edge in the target node at the target iteration number. Figure 4 For example, it is based on Figure 4 The graph structure shown determines the edge (R) of each type for the target node v. 1(v,p) R 1(v,s) R 5(v,w) R 5(v,s)The weighting coefficient at the target iteration number (t) This indicates that the correlation between neighboring nodes p and v at time t is of edge type R1, and so on.

[0142] The weight coefficients used to represent the correlation are related to the target iteration number (i.e., the fixed iteration number) and the edge type. Different edge types and target iteration numbers correspond to different weight coefficients.

[0143] In one possible design, step S301 could be implemented in the following ways:

[0144] Based on the graph structure and feature mapping function clusters, the weight coefficients corresponding to each type of edge of the target node are obtained through differential fusion processing. In other words, based on the graph structure, differential melting is performed using feature mapping function clusters to obtain the weight coefficients corresponding to each type of edge of the target node.

[0145] The feature mapping function cluster includes a set of second feature mapping functions corresponding to each type of edge. Its function is to perform feature dimension transformation. Each feature mapping layer is a neural network with a set of trainable parameters. The set of parameters is related to the target iteration number and the edge type. Different edges have different second feature mapping functions.

[0146] It should be noted that feature dimension transformation specifically refers to converting low-dimensional features into high-dimensional features. A high-dimensional space can better orthogonalize low-dimensional features. Orthogonalized features are independent of each other, better reflecting the principal components of the features, thus increasing the degree of differentiation between data features. This is the essence of differential fusion processing.

[0147] Furthermore, for edges of the same type, the sum of their corresponding weight coefficients is 1, for example... and The sum is 1. and The sum is 1.

[0148] In one possible design, the weight coefficients corresponding to each type of edge of the target node are obtained through differential fusion processing based on the graph structure and feature mapping function clusters. Possible implementation methods include matrix multiplication, proportional scaling, and normalization. The implementation steps are as follows: Figure 6 As shown. Figure 6 This is a flowchart illustrating yet another object recognition method provided in an embodiment of this application. Figure 6 As shown, the embodiments of this application include:

[0149] S401: For each type of edge of the target node, perform matrix multiplication based on the node vectors corresponding to the target node and each neighboring node in the previous iteration number, and the set of second feature mapping functions corresponding to the current type of edge of the target node, and obtain the matrix multiplication result.

[0150] S402: Scale the matrix multiplication result proportionally to obtain the scaled result;

[0151] S403: Normalize the scaling result using the normalized exponential function to obtain the weight coefficients corresponding to each type of edge of the target node.

[0152] As described above, the embodiments of this application obtain the weight coefficients corresponding to each type of edge of the target node through differential fusion processing. The implementation of this differential fusion processing includes matrix multiplication, proportional scaling processing, and normalization.

[0153] The implementation method of differentiated fusion processing is shown in the following formula (7):

[0154]

[0155] In formula (7), information fusion is first performed on the target node (e.g., v) and the neighboring nodes (represented by o) through matrix multiplication, h v and h o Let represent the vector representations of the target node and its neighboring nodes, respectively; both are vectors from the hidden layer. This indicates transposition. Furthermore, through... Proportional scaling is performed to prevent the vector dot product value from the previous matrix multiplication from being too large, which could lead to gradient vanishing or gradient explosion during softmax normalization. In other words, the vector dot product is scaled by proportional scaling. Softmax normalization ensures that the weight coefficients of edges of the same type obtained after applying the softmax function to the input vector are all within the interval (0,1), i.e., S(v,o)∈(0,1).

[0156] In matrix multiplication, a second feature mapping function is introduced as a linear (feature) mapping layer for feature dimension transformation. Based on the introduction of this second feature mapping function, ... Figure 4 Taking the spectrum shown 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) Weighting coefficients The processes are shown in formulas (8) to (11) respectively:

[0157]

[0158]

[0159]

[0160]

[0161] Where Q, K, and L represent the second feature mapping function, for example as well as These are the second feature mapping functions with edge type R1, collectively referred to as the set of second feature mapping functions with edge type R1; as well as These are the second feature mapping functions with edge type R5, collectively referred to as the set of second feature mapping functions with edge type R5. and For edge type R1, this is the node vector corresponding to the neighboring nodes (p and s) of the target node in the previous iteration (t-1). and For edge type R5, this is the node vector corresponding to the neighboring nodes (w and s) of the target node in the previous iteration (t-1). It is the node vector corresponding to the target node v in the previous iteration (t-1).

[0162] Taking formula (8) as an example, the calculation process of formula (8) is described in detail below:

[0163]

[0164] in Received It is a 1×N vector.

[0165]

[0166] in, Received It is an N×1 vector.

[0167] Assumption as well as but

[0168] Due to weighting coefficients It is a real number. It's a proportional scaling process, essentially a hyperparameter that can be preset, for example, set to 8. Let's assume the calculated value is... The 0.88 shown in formula (12) is as follows:

[0169]

[0170] Thus we can obtain

[0171] Similarly, it can be calculated that and

[0172] It should be noted that the values ​​of the weight coefficients mentioned above are only an assumption. In actual working conditions, vector calculations are vector calculations with hundreds of dimensions.

[0173] It is worth noting that, for example, This represents the result of matrix multiplication. This indicates the scaling result.

[0174] Furthermore, L in the second feature mapping function represents the dependency between nodes, used to characterize the correlation calculation between nodes in high-dimensional space. For example, in two-dimensional space, cosine distance is usually used to represent correlation. However, in multi-dimensional space, in order to satisfy the orthogonal coordinate system, L needs to be used for fitting, which is related to the target iteration number and is an N×N weight vector. 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 purpose of feature dimension transformation, and the value of each element is determined by the actual data in the graph structure.

[0175] Thus, through Figure 6 The illustrated embodiment can obtain the weight coefficients corresponding to each type of edge of the target node.

[0176] The object recognition method provided in this application obtains the weight coefficients corresponding to each type of edge of the target node through differential fusion processing, such as matrix multiplication, proportional scaling, and normalization. Specifically, during matrix multiplication, a set of second feature mapping functions is introduced to transform the feature dimension, converting low-dimensionality into high-dimensionality. This makes the orthogonalized features independent, allowing for the selection of principal components in subsequent calculations without needing to calculate all features, thus improving computational accuracy and efficiency. In other words, a neural network with shared parameters across edge types replaces the original transition matrix method with non-shared parameters across edge types. This prevents the model parameters from increasing exponentially with the number of nodes, reducing the number of neural networks required and consequently reducing the computational cost and parameter size of the behavior type prediction model. Furthermore, proportional scaling avoids the problems of gradient explosion and gradient vanishing.

[0177] S302: Determine the feature data transmitted by different types of edges based on the weight coefficients and weight parameters of the network model corresponding to each type of edge.

[0178] After obtaining the weight coefficients corresponding to each type of edge in the target node, the feature data transmitted by different types of edges is further obtained by combining the weight parameter network model.

[0179] For example, in getting Then, by combining the weight parameter network model, different types of edge R are obtained. 1(v,p) R 1(v,s) and R 5(v,w) R 5(v,s) The transmitted feature data and

[0180] The weighted parameter network model includes a first feature mapping function and a feature matrix. The first feature mapping function is only related to the target iteration number; that is, it is a neural network function related to the target iteration number and is independent of the edge type. It is used for feature space transformation. The feature matrix, on the other hand, is related to both the edge type of the target node and the target iteration number. It is used to transform the information dimension when summarizing the isomorphic graph, so as to align the feature dimension of the isomorphic graph with the feature dimension of the summarized graph.

[0181] Assume the first mapping function is V t Let its dimension be M×B, and the feature matrix be... Set its dimensions to B×C.

[0182] In one possible design, step S302 could be implemented in the following ways:

[0183] The product of the weight coefficients for each type of edge, the node vectors of the neighboring nodes for each type of edge before the target iteration number, and the weight parameters of the network model for each type of edge is obtained. This product is then used to determine the feature data passed to the corresponding edge type. The product is represented as a feature vector.

[0184] by Figure 4 For example, R can be calculated using formulas (13) to (16) as shown below. 1(v,p) R 1(v,s) and R 5(v,w) R 5(v,s) The transmitted feature data and

[0185]

[0186]

[0187]

[0188]

[0189] For example, the above Substitute into formula (13) to calculate The process is as follows:

[0190]

[0191] Thus obtained It is a 1×C eigenvector to represent the edge type R. 1(v,p) Transmitted feature data.

[0192] Similarly, we can obtain Let R represent the edge type respectively. 1(v,s) and R 5(v,w) R 5(v,s) Transmitted feature data.

[0193] Thus, based on the weight coefficients and weight parameters corresponding to each type of edge, the network model determines the feature data transmitted by different types of edges, and the feature data is represented by feature vectors.

[0194] As described in the above embodiments, the weight coefficient of each type of edge at the target iteration number is first calculated based on the graph structure. Then, the feature data transmitted by different types of edges is obtained by combining the weight parameter network model. This enables the determination of the feature data transmitted by different types of edges by each neighboring node of the target node based on the graph structure, thus providing a data basis for the summarization of the neighborhood information of the target node, i.e., the determination of the new feature data of the target node at the target iteration number.

[0195] In one possible design, step S202 could be implemented as follows: Figure 7 As shown. Figure 7 This is a flowchart illustrating yet another object recognition method provided in an embodiment of this application. Figure 7 As shown, the embodiments of this application include:

[0196] S501: Obtain the sum of the feature vectors corresponding to each type of edge of the target node, and get the vector sum;

[0197] S502: Apply an activation function to the sum of the vectors to perform piecewise linear processing, resulting in a linear vector;

[0198] S503: Obtain the product of the linear mapping matrix and the linear result vector, and use the obtained product vector to represent the newly added feature data of the target node at the target iteration number.

[0199] The feature vector corresponding to each type of edge is used to characterize the feature data transmitted by each type of edge.

[0200] As described in the previous embodiments, the feature data transmitted by each type of edge is represented by the feature vector corresponding to each type of edge, for example, edge type R 1(v,p) Transmitted feature data This 1×C feature vector representation. First, obtain the sum of the feature vectors corresponding to each type of edge of the target node, and get the vector sum. For example, obtain... The sum of these vectors is then processed using an activation function, such as the ReLU function, to create a piecewise linear sum. The product of this linear mapping matrix and the resulting linear vector is then used to represent the new feature data of the target node at the target iteration number.

[0201] by Figure 4 Taking the feature data of different types of edges in the graph as an example, such as obtaining the sum of vectors, performing piecewise linear processing, and obtaining the product, the process can be achieved through the following formula (17), thus obtaining... Figure 4 The newly added feature data of target node v at the target iteration number (t) is shown.

[0202]

[0203] Among them, M t This represents a linear mapping matrix, which is a neural network in the fusion layer, primarily used to align vector dimensions; for example, it can be a C×M vector. It can be understood that the dimension of the linear mapping matrix is ​​determined by the dimension of the feature vectors corresponding to different types of edges, while the specific values ​​of each element in the linear mapping matrix are determined by the actual data of the graph structure.

[0204] Activation functions, such as the ReLU function, are used to increase the non-linearity of neural networks. By introducing non-linear factors, the model's ability to withstand data noise is increased.

[0205] When running formula (17), the linear result vector obtained by piecewise linear processing using the activation function is shown in formula (18) below:

[0206]

[0207] Among them, for the linear result vector In other words, with For example, its operation satisfies the following formula (19) to retain negative samples, that is, to introduce nonlinear samples.

[0208]

[0209] This is an adjustable hyperparameter, for example, it can be set to 0.01. As can be seen from formula (19), for... For any of them, when it is less than or equal to 0, the corresponding linear result vector is not directly set to 0 to delete negative samples, but is instead determined by introducing a hyperparameter. The exponent keeps the linear result vector within a certain range, which is beneficial to the recognition accuracy of the subsequent behavior type prediction model when faced with various noises.

[0210] Substituting formula (18) and the linear mapping matrix into formula (17), we obtain the following formula (20):

[0211]

[0212] Therefore, the resulting product vector is used to represent the newly added feature data of the target node at the target iteration number.

[0213] Thus, by obtaining the sum of vectors, performing piecewise linear processing, and obtaining the product, the new feature data of the target node at the target iteration number is determined based on the feature data passed by different types of edges.

[0214] The object recognition method provided in this application, after obtaining the feature data transmitted by different types of edges of the target node, obtains the newly added feature data of the target node at the target iteration number by acquiring the sum of feature vectors, performing piecewise linear processing, and obtaining the product. That is, at the time corresponding to the target iteration number, the graph structure propagates the newly added information. The piecewise linear processing uses an activation function, which can retain negative samples for subsequent calculations, thereby increasing the robustness of the behavior type prediction model to data noise and improving the recognition accuracy.

[0215] In one possible design, step S203 could be implemented in the following ways:

[0216] The product vector and the target node's hidden layer vector at the previous iteration number are input into a predefined recurrent neural network, and the output is determined as a node vector. For example, through... and Calculate the hidden layer vector of the node at time t corresponding to the target iteration number. That is, the node vector of the target node.

[0217] As described in the foregoing embodiments, for the target node, at time t, the target node's hidden layer... The feature data to be expressed by the vector, that is, the node vector, comes from two parts: one part is newly added feature data, that is... The other part comes from the hidden layer vector encoding of the target node itself at the previous time step (i.e., t-1). The former refers to the neighborhood information learned at time t, while the latter is the necessary input for the temporal features.

[0218] Based on the above two parts, node vectors can be obtained through a pre-defined recurrent neural network. This pre-defined recurrent neural network can be, for example, a gated recurrent unit (GRU), a type of neural network used to capture dependencies in time-series data. It can preserve information in long sequences and filter out irrelevant information for prediction. The implementation principle of GRU is not elaborated in this application.

[0219] The implementation of new feature data represented by product vectors and the acquisition of node vectors by a pre-defined recurrent neural network such as GRU can be achieved by the following formula (21):

[0220]

[0221] Among them, a pre-defined recurrent neural network, such as a GRU, receives... and Used for updating GRU output It is a 1×M feature vector, and the initial vector of the target node. Same dimensions.

[0222] After the target number of iterations, for example, after 5 iterations, then take... As the final vector representation of the target node, that is, the node vector of the target node, it is then used... For example Identify objects with distinctive behaviors. The node vectors are used to characterize the behavior type prediction model.

[0223] The object recognition method provided in this application, based on the newly added feature data represented by the product vector and the hidden layer vector of the target node at the previous iteration number, obtains the node vector of the target node, i.e., the behavior type prediction model, through a preset recurrent neural network. The preset recurrent neural network, because it can preserve information from long sequences and filter out information irrelevant to the prediction, can discard some outdated features and select features more relevant to the current time. This improves the computational accuracy of the obtained behavior type prediction model, enabling accurate and efficient feature behavior risk identification of the object to be identified.

[0224] Figure 8 This is a schematic diagram of the structure of an object recognition device provided in an embodiment of this application. Figure 8 As shown, the object recognition device 600 provided in this application embodiment includes:

[0225] The first processing module 601 is used to acquire the portrait feature data and correlation of the object to be identified, and generate the map structure of the object to be identified based on the portrait feature data and correlation. The object to be identified includes enterprise objects or individual objects.

[0226] The second processing module 602 is used to obtain the feature behavior probability distribution of the object to be identified based on the graph structure and behavior type prediction model. The behavior type prediction model is obtained by training the model using the parameter sharing neighborhood information fusion mechanism of edge type and graph structure.

[0227] The third processing module 603 is used to determine whether the object to be identified is a characteristic behavior object based on the characteristic behavior probability distribution, and whether the characteristic behavior object has characteristic behavior risk.

[0228] exist Figure 8 On this basis, Figure 9 This is a schematic diagram of another object recognition device provided in an embodiment of this application. Figure 9 As shown, the object recognition device 600 provided in this embodiment of the application further includes: a model learning module 604. The model learning module 604 is used for:

[0229] The target feature data of the target node is determined based on the graph structure. The target feature data includes the feature data transmitted by each neighboring node of the target node through different types of edges. The target node is the node in the graph structure that represents the object to be identified.

[0230] Based on the feature data passed by each type of edge, determine the new feature data of the target node at the target iteration number;

[0231] The node vector is obtained by combining the newly added feature data and the target node's own hidden layer vector at the previous target iteration number. The hidden layer vector is used to represent the feature data that the target node self-fused to at the previous target iteration number, and the node vector is used to represent the behavior type prediction model.

[0232] In one possible design, the model learning module 604 is also used for:

[0233] The weight coefficients of each type of edge of the target node at the target iteration number are determined based on the graph structure. The weight coefficients are used to characterize the correlation between the target node and each neighboring node.

[0234] Based on the weight coefficients and weight parameters corresponding to each type of edge, the network model determines the feature data transmitted by different types of edges.

[0235] The weighted parameter network model includes a first feature mapping function and a feature matrix. The first feature mapping function is related to the target iteration number and is used for feature space transformation. The feature matrix is ​​related to the edge type of the target node and the target iteration number.

[0236] In one possible design, the model learning module 604 is also used for:

[0237] Based on the graph structure and feature mapping function cluster, the weight coefficients corresponding to each type of edge of the target node are obtained through differential fusion processing;

[0238] In this context, 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 edge.

[0239] In one possible design, the model learning module 604 is also used for:

[0240] For each type of edge of the target node, matrix multiplication is performed based on the node vectors corresponding to the target node and each neighboring node in the previous iteration number, as well as the set of second feature mapping functions corresponding to the edge of the current type of the target node, to obtain the matrix multiplication result.

[0241] The matrix multiplication result is scaled proportionally to obtain the scaled result.

[0242] The scaling result is normalized using a normalized exponential function to obtain the weight coefficients for each type of edge of the target node.

[0243] In one possible design, the model learning module 604 is also used for:

[0244] The product of the weight coefficients corresponding to each type of edge, the node vector of the neighboring nodes of each type of edge before the target iteration number, and the weight parameters of each type of edge in the network model is obtained, and the resulting product is determined as the feature data transmitted by the edge of the corresponding type.

[0245] In one possible design, the model learning module 604 is also used for:

[0246] Obtain the sum of the feature vectors corresponding to each type of edge of the target node, and the feature vector corresponding to each type of edge is used to characterize the feature data transmitted by each type of edge;

[0247] The sum of the vectors is processed piecewise linearly using an activation function to obtain a linear result vector;

[0248] Obtain the product of the linear mapping matrix and the linear result vector, and use the resulting product vector to represent the newly added feature data of the target node at the target iteration number.

[0249] In one possible design, the model learning module 604 is also used for:

[0250] The product vector and the target node's own hidden layer vector at the previous iteration number are input into a preset recurrent neural network, and the output is determined as a node vector.

[0251] In one possible design, the second processing module 602 is also used for:

[0252] The node vector is mapped to a class probability distribution vector through the third feature mapping function;

[0253] The category probability distribution vector is input into the binary classifier to obtain the category output vector, which is used to characterize the probability distribution of the feature behavior of the object to be identified.

[0254] In one possible design, the third processing module 603 is also used for:

[0255] The system determines whether the object to be identified is a characteristic behavior object based on the category output vector and the preset category vector. The preset category vector includes the category vector of characteristic behavior objects and the category vector of non-characteristic behavior objects.

[0256] In one possible design, the first processing module 601 is further used for:

[0257] Based on the profile feature data of the object to be identified, identify the related objects that have transactional relationships with the object to be identified;

[0258] Determine the relationship categories between each associated object and the object to be identified based on the transaction data corresponding to the transactions;

[0259] The association relationships of the object to be identified are obtained based on the relationship category.

[0260] In one possible design, the first processing module 601 is further used for:

[0261] Set the objects to be identified as nodes in the graph structure, and set the edges of the graph structure according to the relationships of the objects to be identified, so as to generate the graph structure of the objects to be identified.

[0262] In one possible design, the first processing module 601 is further used for:

[0263] The eigenvalued portrait feature data of the object to be identified is used as the initial vector of the target node in the graph structure.

[0264] The object recognition device provided in this application embodiment can execute each step of the object recognition method in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0265] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 10 As shown, the electronic device 700 may include a processor 701 and a memory 702 communicatively connected to the processor 701.

[0266] The memory 702 is used to store programs. Specifically, the program may include program code, which includes computer-executable instructions.

[0267] The memory 702 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0268] The processor 701 is used to execute computer execution instructions stored in the memory 702 to implement the object recognition method.

[0269] 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 this application.

[0270] Optionally, the memory 702 can be either standalone or integrated with the processor 701. When the memory 702 is a device independent of the processor 701, the electronic device 700 may further include:

[0271] Bus 703 is used to connect processor 701 and memory 702. The bus can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus, or an extended industry standard architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not mean there is only one bus or one type of bus.

[0272] Optionally, in a specific implementation, if the memory 702 and the processor 701 are integrated on a single chip, the memory 702 and the processor 701 can communicate through an internal interface.

[0273] This application also provides a computer-readable storage medium, which may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk. Specifically, the computer-readable storage medium stores computer-executable instructions, which are used in the various steps of the methods described in the above embodiments.

[0274] This application also provides a computer program product, including computer execution instructions that, when executed by a processor, implement the steps of the methods described above.

[0275] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0276] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An object recognition method, characterized in that, include: Acquire the portrait feature data and association relationships of the object to be identified, and generate the map structure of the object to be identified based on the portrait feature data and the association relationships. The object to be identified includes enterprise objects or individual objects. Based on the graph structure and behavior type prediction model, the feature behavior probability distribution of the object to be identified is obtained. The behavior type prediction model is learned by learning the graph structure using a parameter-sharing neighborhood information fusion mechanism of edge type. Based on the probability distribution of the characteristic behaviors, it is determined whether the object to be identified is a characteristic behavior object, and the characteristic behavior object has characteristic behavior risk; The method of learning the graph structure using the parameter-sharing neighborhood information fusion mechanism of the edge type to obtain the behavior type prediction model includes: The target feature data of the target node is determined based on the graph structure. The target feature data includes feature data transmitted by each neighboring node of the target node through different types of edges. The target node is the node in the graph structure that represents the object to be identified. Based on the feature data transmitted by different types of edges, determine the new feature data of the target node at the target iteration number; The node vector is obtained based on the newly added feature data and the target node's own hidden layer vector at the previous target iteration number. The self-hidden layer vector is used to represent the feature data that the target node self-fused to at the previous target iteration number. The node vector is used to represent the behavior type prediction model. The step of determining the target feature data of the target node based on the graph structure includes: The weight coefficients of each type of edge of the target node are determined according to the graph structure at the target iteration number, and the weight coefficients are used to characterize the correlation between the target node and each of the neighboring nodes; The feature data transmitted by each type of edge is determined based on the weight coefficients and weight parameter network model corresponding to each type of edge. The weighted parameter network model includes a first feature mapping function and a feature matrix. The first feature mapping function is related to the target iteration number and is used for feature space transformation. The feature matrix is ​​related to the edge type of the target node and the target iteration number.

2. The object recognition method according to claim 1, characterized in that, The step of determining the weight coefficient of each type of edge of the target node at the target iteration number based on the graph structure includes: Based on the graph structure and the feature mapping function cluster, the weight coefficients corresponding to each type of edge of the target node are obtained through differential fusion processing; Wherein, 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 edge.

3. The object recognition method according to claim 2, characterized in that, The step of obtaining the weight coefficients for each type of edge of the target node through differential fusion processing based on the graph structure and the set of feature mapping functions includes: For each type of edge of the target node, matrix multiplication is performed based on the node vectors corresponding to the target node and each of the neighboring nodes in the previous iteration of the target iteration, and the set of second feature mapping functions corresponding to the edge of the current type of the target node, to obtain the matrix multiplication result. The matrix multiplication result is scaled proportionally to obtain the scaled result. The scaling result is normalized using a normalized exponential function to obtain the weight coefficients corresponding to each type of edge of the target node.

4. The object recognition method according to claim 1, characterized in that, The step of determining the feature data transmitted by each type of edge based on the weight coefficients and weight parameter network model corresponding to each type of edge includes: The product of the weight coefficients corresponding to each type of edge, the node vectors of the neighboring nodes corresponding to each type of edge before the target iteration number, and the weight parameters of the network model corresponding to each type of edge is obtained, and the resulting product is determined as the feature data transmitted by the edge of the corresponding type.

5. The object recognition method according to any one of claims 1-4, characterized in that, The step of determining the new feature data of the target node at the target iteration number based on the feature data passed by each type of edge includes: The sum of feature vectors corresponding to each type of edge of the target node is obtained, and the feature vector corresponding to each type of edge is used to characterize the feature data transmitted by each type of edge. The sum of the vectors is processed piecewise linearly using an activation function to obtain a linear result vector; Obtain the product of the linear mapping matrix and the linear result vector, and use the resulting product vector to represent the newly added feature data of the target node at the target iteration number.

6. The object recognition method according to claim 5, characterized in that, The step of obtaining the node vector based on the newly added feature data and the target node's own hidden layer vector at the previous target iteration number includes: The product vector and the hidden layer vector of the target node at the previous iteration number are input into a preset recurrent neural network, and the output is determined as the node vector.

7. The object recognition method according to any one of claims 1-4, characterized in that, The step of obtaining the probability distribution of the object to be identified based on the graph structure and behavior type prediction model includes: The node vector is mapped to a category probability distribution vector using a third feature mapping function; The category probability distribution vector is input into a binary classifier to obtain a category output vector, which is used to characterize the feature behavior probability distribution of the object to be identified.

8. The object recognition method according to claim 7, characterized in that, Determining whether the object to be identified is a characteristic behavior object based on the characteristic behavior probability distribution includes: The object to be identified is determined based on the category output vector and the preset category vector. The preset category vector includes the category vector of the characteristic behavior object and the category vector of non-characteristic behavior objects.

9. The object recognition method according to any one of claims 1-4, characterized in that, Obtaining the association relationship of the object to be identified includes: Based on the profile feature data of the object to be identified, obtain the associated objects that have transactional relationships with the object to be identified; The relationship category between each associated object and the object to be identified is determined based on the transaction data corresponding to the transactions. The association relationship of the object to be identified is obtained based on the relationship category.

10. The object recognition method according to claim 9, characterized in that, The step of generating the atlas structure of the object to be identified based on the portrait feature data and the correlation includes: The object to be identified is set as a node of the graph structure, and the edges of the graph structure are set according to the association relationship of the object to be identified, so as to generate the graph structure of the object to be identified.

11. The object recognition method according to any one of claims 1-4, characterized in that, After obtaining the portrait feature data of the object to be identified, the method further includes: The eigenvalued portrait feature data of the object to be identified is used as the initial vector of the target node in the graph structure.

12. An object recognition device, characterized in that, The object recognition device is configured to perform the object recognition method according to any one of claims 1 to 11, the object recognition device comprising: The first processing module is used to acquire the portrait feature data and association relationships of the object to be identified, and generate the map structure of the object to be identified based on the portrait feature data and the association relationships. The object to be identified includes enterprise objects or individual objects. The second processing module is used to obtain the feature behavior probability distribution of the object to be identified based on the graph structure and behavior type prediction model. The behavior type prediction model is obtained by learning the graph structure using a parameter-sharing neighborhood information fusion mechanism of edge type. The third processing module is used to determine whether the object to be identified is a characteristic behavior object based on the characteristic behavior probability distribution, and the characteristic behavior object has characteristic behavior risk.

13. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the object recognition method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the object recognition method as described in any one of claims 1 to 11.

15. A computer program product comprising computer-executable instructions, characterized in that, When the computer execution instructions are executed by the processor, they are used to implement the object recognition method as described in any one of claims 1 to 11.

Images