Object representation method and apparatus, and device and computer-readable storage medium

Abstract

Disclosed in the present application are an object representation method and apparatus, and a device and a computer-readable storage medium. The object representation method comprises: on the basis of the association relationships between a plurality of objects, constructing a topological structure in which the plurality of objects are used as nodes and the association relationships are used as connecting edges; inputting the topological structure and an initial node feature matrix into an object representation model; using each of a plurality of graph neural network layers in the object representation model to extract common information and differential information of neighbor nodes of each node in the topological structure and use same as feature information of the node, and outputting node feature matrices that respectively correspond to the plurality of graph neural network layers; and performing first aggregation processing on the initial node feature matrix and the node feature matrices that respectively correspond to the plurality of graph neural network layers, so as to generate pieces of object representation information that respectively correspond to the plurality of objects.

Description

Object representation method, device, apparatus, and computer-readable storage medium

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Chinese patent application 202311667924.2, filed on December 6, 2023, entitled “Object representation method, apparatus, device and computer-readable storage medium,” the entire contents of which are incorporated herein by reference. Technical Field

[0003] The present application relates to artificial intelligence technology, and more particularly to an object representation method, apparatus, device, and computer-readable storage medium. Background Art

[0004] With the development of online shopping, social networking, and video platforms, recommendation systems have become an indispensable part of online platforms. Accurately extracting the representation information of each recommended object and the recommended object in a recommendation system is an important prerequisite for improving the recommendation effect.

[0005] Currently, when using a neural network model to represent each recommended object and the recommended object, the features of each object often have saturated values, that is, the problem of over-smoothing occurs, thereby reducing the accuracy of object representation. Summary of the Invention

[0006] The embodiments of the present application provide an object characterization method, apparatus, device, and computer-readable storage medium, which can effectively improve the over-smoothing problem during object characterization, thereby improving the accuracy of object characterization.

[0007] In a first aspect, an embodiment of the present application provides an object characterization method, the method comprising:

[0008] Get the relationship between multiple objects;

[0009] Constructing a topological structure with the plurality of objects as nodes according to the association relationship, wherein nodes corresponding to the objects having the association relationship in the topological structure have connecting edges;

[0010] Inputting the topological structure and the initial node feature matrix into an object representation model, wherein the object representation model includes multiple graph neural network layers, and the initial node feature matrix includes initial feature information corresponding to each node in the topological structure;

[0011] Utilizing each of the multiple graph neural network layers, for each node in the topological structure, extracting common information and difference information of its neighboring nodes as feature information of the node, and outputting node feature matrices corresponding to the multiple graph neural network layers respectively;

[0012] A first aggregation process is performed on the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers to generate object representation information corresponding to the multiple objects.

[0013] In a second aspect, an embodiment of the present application provides an object representation device, the device comprising:

[0014] Relationship acquisition module, used to obtain the association relationship between multiple objects;

[0015] A topology construction module, configured to construct a topology structure with the plurality of objects as nodes according to the association relationship, wherein the nodes corresponding to the objects having the association relationship in the topology structure have connecting edges;

[0016] An information input module, configured to input the topological structure and the initial node feature matrix into an object representation model, wherein the object representation model includes a plurality of graph neural network layers, and the initial node feature matrix includes initial feature information corresponding to each node in the topological structure;

[0017] An information extraction module is configured to utilize each of the multiple graph neural network layers to extract, for each node in the topological structure, common information and difference information of its neighboring nodes as feature information of the node, and output a node feature matrix corresponding to each of the multiple graph neural network layers;

[0018] A representation generation module is used to perform a first aggregation process on the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers, so as to generate object representation information corresponding to the multiple objects.

[0019] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising: a processor and a memory storing computer program instructions;

[0020] When the processor executes the computer program instructions, the steps of the object characterization method as described in any one of the embodiments of the first aspect are implemented.

[0021] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, the steps of the object characterization method as described in any one of the embodiments of the first aspect are implemented.

[0022] In a fifth aspect, an embodiment of the present application provides a computer program product. When the instructions in the computer program product are executed by a processor of an electronic device, the electronic device performs the steps of the object characterization method as described in any embodiment of the first aspect.

[0023] The object representation method, apparatus, device and computer-readable storage medium in the embodiments of the present application, by utilizing the association relationship between multiple objects to construct a topological structure with objects as nodes and association relationships as connecting edges, and utilizing each graph neural network layer in the object representation model, extracts the common information and difference information of its neighboring nodes for each node in the topological structure as the feature information of the node, and then performs a first aggregation process on the node feature matrix output by the object representation model and corresponding to the multiple graph neural network layers contained therein, and finally generates object representation information corresponding to the multiple objects. In this way, since the embodiment of the present application extracts the common information and difference information of the neighboring nodes corresponding to each node, not only the common information between the nodes but also the difference information between different nodes is retained, it can effectively improve the over-smoothing problem during object representation, thereby improving the accuracy of object representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0025] FIG1 is a schematic diagram of a flow chart of an object characterization method provided by one embodiment of the present application;

[0026] FIG2 is a schematic diagram of the structure of an object representation model provided by this application;

[0027] FIG3 is a schematic structural diagram of an object characterization device provided by one embodiment of the present application;

[0028] FIG4 is a schematic structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0029] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present application by illustrating the examples of the present application.

[0030] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, the elements defined by the phrase "comprising..." do not exclude the presence of other identical elements in the process, method, article, or device comprising the elements.

[0031] The acquisition, storage, use, and processing of data (including but not limited to the features and information herein) in the technical solution of this application comply with the relevant provisions of national laws and regulations.

[0032] In recommendation systems, taking e-commerce platforms as an example, most platforms contain a large amount of implicit information, such as users' browsing, collection, and purchase records. Collaborative filtering can extract the similarities between users and products in the implicit information, thereby recommending products that users may be interested in.

[0033] In recent years, with the development of artificial intelligence (AI), neural network models have begun to be applied to collaborative filtering, and neural network-based collaborative filtering methods have been proposed. In theory, increasing the number of neural network layers in a model allows it to capture more detailed information about each object, thereby improving the accuracy of the recommendation system. However, each neural network layer attenuates the differential information between different objects, leading to saturation of object features. This problem is known as oversmoothing. In practice, it has been found that when the number of neural network layers is large, the accuracy of the recommendation system tends to saturate or even decreases. This is the manifestation of the oversmoothing problem in collaborative filtering.

[0034] Through theoretical analysis, it can be seen that when the number of neural network layers in the existing collaborative filtering method based on neural networks is large, the feature information of each object node will be over-filtered, thereby losing the individual difference information between nodes, which will lead to the over-smoothing problem and reduce the accuracy of object representation.

[0035] To address the problems of the prior art, embodiments of the present application provide an object characterization method, apparatus, device, and computer-readable storage medium. The object characterization method can be applied to scenarios where recommended objects or recommended objects are characterized. The object characterization method provided in embodiments of the present application is first introduced below.

[0036] Figure 1 is a flow chart of an object characterization method provided by one embodiment of the present application. The object characterization method can be executed by electronic devices such as computers and servers.

[0037] As shown in FIG1 , the object representation method may specifically include the following steps:

[0038] S110, obtaining association relationships between multiple objects;

[0039] S120: constructing a topological structure with multiple objects as nodes based on the association relationship, wherein nodes corresponding to the objects with the association relationship in the topological structure have connecting edges;

[0040] S130, inputting the topological structure and the initial node feature matrix into an object representation model, where the object representation model includes multiple graph neural network layers, and the initial node feature matrix includes initial feature information corresponding to each node in the topological structure;

[0041] S140. Utilizing each of the multiple graph neural network layers, for each node in the topological structure, extract commonality information and difference information of its neighboring nodes as feature information of the node, and outputting node feature matrices corresponding to the multiple graph neural network layers, respectively.

[0042] S150. Perform a first aggregation process on the initial node feature matrix and the node feature matrices corresponding to multiple graph neural network layers to generate object representation information corresponding to multiple objects.

[0043] Thus, by utilizing the association relationship between multiple objects to construct a topological structure with objects as nodes and association relationships as connecting edges, and utilizing each graph neural network layer in the object representation model, the common information and difference information of the neighboring nodes of each node in the topological structure are extracted as the feature information of the node, and then the node feature matrix output by the object representation model corresponding to the multiple graph neural network layers contained therein is subjected to a first aggregation process, and finally the object representation information corresponding to the multiple objects is generated. In this way, since the common information and difference information of the neighboring nodes corresponding to each node are extracted in the embodiment of the present application, not only the common information between the nodes but also the difference information between different nodes is retained, the over-smoothing problem during object representation can be effectively improved, thereby improving the accuracy of object representation.

[0044] The specific implementation methods of the above steps are introduced below.

[0045] In some embodiments, in S110 , the objects may include recommending objects and recommended objects in the recommendation system, such as users and products, where products may include items or information to be recommended, etc. In addition, the association relationship between objects may be determined through interaction information between the objects.

[0046] For example, taking users and products as an example, the association relationships between users and users, users and products, and products and products can be mined from massive interactive information. The interactive information may include, for example, user feedback, ratings and comments on products, user clicks on events and links, user browsing or purchasing behavior on products, etc.

[0047] In some embodiments, in S120, after the association relationships between the multiple objects are obtained, a topological structure G may be constructed based on the association relationships, with the multiple objects as nodes and the association relationships between the objects as connecting edges. The topological structure G may be a connected graph, meaning that there are no isolated nodes in the topological structure G. Furthermore, a node in the topological structure G may correspond to an object, and connecting edges may exist between nodes corresponding to objects having association relationships.

[0048] For example, if user A has browsed product B, there is an association relationship between user A and product B. Then, when constructing the topological structure G, a connection edge can be established between the node corresponding to user A and the node corresponding to product B.

[0049] In some embodiments, in S130 , the object representation model may be a neural network model for representing an object, and the object representation model may include multiple graph neural network layers.

[0050] Here, before inputting the topological structure G into the object representation model, the topological structure G can be converted into an adjacency matrix representation. If there are N nodes in the topological structure, an N*N adjacency matrix W can be generated. In practice, to simplify computational complexity, the complete adjacency matrix W can be generated directly from the topological structure G. In addition, to prevent overfitting and improve the generalization ability of the model, some nodes and edges in the topological structure G can be randomly discarded during model training. In this case, the adjacency matrix W needs to change dynamically as nodes and edges are discarded.

[0051] In addition, the initial node feature matrix E 0 It can be a learnable parameter matrix constructed for the nodes corresponding to each object. The initial node feature matrix E 0 It can be determined after the model training is completed. If there are N nodes in the topological structure, and the initial feature information corresponding to each node has d feature dimensions preset, then the initial node feature matrix E 0 It can be an N*d matrix.

[0052] It should be noted that before model training, the initial node feature matrix E 0 The initial feature information corresponding to each node included in can be some preset information. During the model training process, the initial node feature matrix E can be continuously adjusted. 0 The initial feature information corresponding to each node in the , and after the model training is completed, the final initial node feature matrix E is determined 0 .

[0053] For example, the adjacency matrix W generated by the topological structure G can be compared with the initial node feature matrix E 0 As the input features of the object representation model, it is input into the object representation model so that the object representation model can perform corresponding information processing.

[0054] In some embodiments, in S140, since the object representation model includes multiple graph neural network layers, each graph neural network layer may process the input information in a similar manner. That is, each layer extracts common information and difference information from the neighboring nodes corresponding to each node in the topological structure G as feature information of the node. Common information may refer to feature information shared by similar nodes, and difference information may refer to feature information that is not shared between nodes.

[0055] In this way, each graph neural network layer can output a node feature matrix, and then multiple node feature matrices corresponding to the multiple graph neural network layers can be obtained.

[0056] In addition, it should be noted that the node feature matrix output by each graph neural network layer can also be used as the input of the next graph neural network layer, so that the next graph neural network layer can continue to extract features based on the output results of the previous layer, thereby improving the accuracy of feature extraction.

[0057] In some embodiments, in S150, after obtaining multiple node feature matrices output by multiple graph neural network layers, the multiple node feature matrices can be compared with the initial node feature matrix E 0 Perform the first aggregation process to combine the multiple node feature matrices with the initial node feature matrix E 0 Aggregated into a matrix. The feature information of each node contained in the matrix can be used as the object representation information of the corresponding object, and the object representation information can be used to represent the characteristics of the object.

[0058] Based on this, when making product recommendations, the similarity between the user and different products can be calculated based on the object representation information, and then products with a close similarity to the user's object representation information can be recommended to the user.

[0059] In addition, in some implementations, the above S130 may specifically include:

[0060] Based on the initial node feature matrix, determining a target node feature matrix to be input into a target graph neural network layer, where the target graph neural network layer is any one of the multiple graph neural network layers;

[0061] Input the topology structure and target node feature matrix into the target graph neural network layer;

[0062] The target graph neural network layer is used to extract the common information and difference information corresponding to the neighboring nodes of each node in the topological structure as the feature information of the node, and the node feature matrix corresponding to the target graph neural network layer is output.

[0063] Here, the node feature matrices corresponding to the inputs of different graph neural network layers in the object representation model can be different.

[0064] For example, for any graph neural network layer, for example, the nth graph neural network layer, where n is an integer greater than 0, the initial node feature matrix E 0 Based on this, the target node feature matrix to be input to the nth graph neural network layer is determined, and then the target node feature matrix and the adjacency matrix W generated according to the topological structure G are used as the input features of the nth graph neural network layer and input into the nth graph neural network layer. The common information and difference information of the neighboring nodes corresponding to each node are extracted by using the nth graph neural network layer as the feature information of the node, and then the node feature matrix E corresponding to the nth graph neural network layer can be output. n .

[0065] Based on this, in some embodiments, determining the target node feature matrix to be input to the target graph neural network layer based on the initial node feature matrix may specifically include:

[0066] When the target graph neural network layer is the first graph neural network layer among multiple graph neural network layers, the initial node feature matrix is ​​used as the target node feature matrix to be input into the target graph neural network layer;

[0067] When the target graph neural network layer is any graph neural network layer except the first graph neural network layer among multiple graph neural network layers, the node feature matrix corresponding to the previous graph neural network layer of the target graph neural network layer is used as the target node feature matrix to be input into the target graph neural network layer.

[0068] Here, multiple graph neural network layers in the object representation model can be arranged sequentially.

[0069] Based on this, in some examples, for the first graph neural network layer in the object representation model, the target node feature matrix to be input to the first graph neural network layer can be the initial node feature matrix E 0 In other examples, for subsequent graph neural network layers other than the first graph neural network layer in the object representation model, for example, the nth graph neural network layer, where n is an integer greater than 1, the target node feature matrix to be input to the nth graph neural network layer can be the node feature matrix E output by the n-1th graph neural network layer. n-1 .

[0070] In addition, in some embodiments, the node feature matrix corresponding to the previous graph neural network layer of the target graph neural network layer is used as the target node feature matrix to be input to the target graph neural network layer, including:

[0071] For each feature dimension in the node feature matrix corresponding to the previous graph neural network layer, the node feature matrix is ​​normalized to obtain the standard node feature matrix corresponding to the previous graph neural network layer;

[0072] The standard node feature matrix is ​​used as the target node feature matrix to be input into the target graph neural network layer.

[0073] Here, if the node feature matrix corresponding to the previous graph neural network layer is an N*d matrix, where N is the number of nodes and d is the feature dimension, the normalization process can be a vertical normalization (batch norm) process.

[0074] For example, the normalization process may be performed according to the following formula (1).

[0075] Among them, E n is the node feature matrix before normalization, is the standard node feature matrix after standardization, and batch_norm() is the vertical normalization function.

[0076] The purpose of the standardization processing in the embodiment of the present application can be to map the eigenvalues ​​under each feature dimension to a feature value range centered on 0, so as to make the difference in the eigenvalues ​​between two dissimilar nodes larger. For example, for two dissimilar nodes, the eigenvalues ​​of a certain dimension are all positive before the standardization processing. After the standardization processing, the eigenvalue of one node can be mapped to a positive value and the eigenvalue of one node can be mapped to a negative value, thereby expanding the difference in the eigenvalues ​​between the two nodes.

[0077] In this way, normalization can amplify the differences between nodes while eliminating the offset in similarity between nodes, further reducing the impact of oversmoothing. In addition, it can also accelerate the convergence process of the model during training.

[0078] It should be noted that for the initial node feature matrix E to be input into the first graph neural network layer 0 , and its corresponding standard node feature matrix It can be itself, that is,

[0079] In addition, in some embodiments, the target graph neural network layer is used to extract the common information and difference information corresponding to the neighboring nodes of each node in the topological structure as the feature information of the node, and the node feature matrix corresponding to the target graph neural network layer is output, including:

[0080] Using the target graph neural network layer, determine the t neighbor nodes corresponding to the target node from the topological structure. The target node is any node in the topological structure.

[0081] Obtain the feature information corresponding to the target node and the feature information corresponding to t neighboring nodes from the target node feature matrix;

[0082] Calculate the similarity between the target node and the t neighbor nodes based on the feature information corresponding to the target node and the feature information corresponding to the t neighbor nodes;

[0083] Get the k neighbor nodes with the highest similarity to the target node and the s neighbor nodes with the lowest similarity to the target node from the t neighbor nodes;

[0084] Extract common information from the feature information corresponding to k neighbor nodes, and extract difference information from the feature information corresponding to s neighbor nodes;

[0085] Determine target feature information corresponding to the target node based on the commonality information and the difference information, and include the target feature information corresponding to the target node in the node feature matrix corresponding to the target graph neural network layer;

[0086] Wherein, t, k, and s are all integers greater than 1, and t≥k+s.

[0087] For example, the neighbor nodes corresponding to each node can be determined through the topological structure G. Taking the target graph neural network layer as the nth graph neural network layer and the target node as node u as an example, the node feature matrix input to the nth graph neural network layer can be obtained from Get the feature information corresponding to node u And the feature information corresponding to its t neighbor nodes

[0088] Based on this, when calculating the similarity between node u and each neighboring node, the following formula (2) can be used to calculate the similarity between nodes. Taking the calculation of the similarity between node i and node j as an example, it is shown below.

[0089] in, represents the similarity between node i and node j in the nth graph neural network layer; sim() is a similarity function, which can be, for example, a function for calculating vector product or cosine similarity; and The node feature matrix of node i and node j at the input is The corresponding feature information.

[0090] For node u, we can select k neighbor nodes with the greatest similarity and s neighbor nodes with the least similarity, where k can be equal to s. The k neighbor nodes are regarded as the nodes with the greatest similarity to node u, and the s neighbor nodes are regarded as the nodes with the least similarity to node u. The feature information corresponding to the node with the greatest similarity includes, for example: The feature information corresponding to the minimum similarity node includes, for example: Feature information corresponding to the k neighbor nodes with the greatest similarity Can be used to extract common information Feature information corresponding to the k neighbor nodes with the smallest similarity Can be used to extract difference information

[0091] Using this common information and difference information The new feature information corresponding to the node u can be finally determined Based on the new feature information corresponding to each node, the node feature matrix E corresponding to the nth graph neural network layer can be combined n , and then output the node feature matrix E n .

[0092] Based on this, in some embodiments, extracting common information from the feature information corresponding to the k neighboring nodes may specifically include:

[0093] According to the similarity between the target node and the k neighbor nodes, the influence weight of each neighbor node on the target node is determined;

[0094] Based on the influence weight of each neighbor node on the target node among the k neighbor nodes, the feature information corresponding to the k neighbor nodes is weighted to obtain common information.

[0095] For example, taking the target graph neural network layer as the nth graph neural network layer and the target node as node u, weighted processing can be performed according to the following formula (3) to calculate the common information.

[0096] in, Represents the common information of node u (subscript l represents low frequency, superscript n represents the nth graph neural network layer); represents the maximum similarity node set of node u, that is, the set of k neighbor nodes with the greatest similarity to node u; α u→i Represents the influence weight of neighbor node i on node u.

[0097] Here, the influence weight of each neighbor node on the target node can be determined based on the similarity between the target node and the k neighbor nodes, for example, obtained by normalizing the similarity.

[0098] Based on this, in some embodiments, determining the influence weight of each of the k neighbor nodes on the target node based on the similarity between the target node and the k neighbor nodes may specifically include:

[0099] The similarity between the target node and k neighbor nodes is normalized to obtain the influence weights of the k neighbor nodes on the target node.

[0100] Here, the normalization method includes but is not limited to using a preset activation function for processing. The preset activation function may be, for example, a softmax function.

[0101] For example, taking the target graph neural network layer as the nth graph neural network layer and the target node as node u, the influence weights of the k neighboring nodes on node u can be determined according to the following formula (4).

[0102] in, represents the similarity between node u and its neighbor node i, Represents the similarity between neighbor node j and neighbor node i in the maximum similarity node set of node u.

[0103] In addition, in order to further improve the accuracy of common information extraction and thus improve the recommendation effect of the recommendation system, smoothing can be performed during normalization. Based on this, in some embodiments, the above-mentioned normalization of the similarity between the target node and the k neighboring nodes to obtain the influence weights of the k neighboring nodes on the target node can specifically include:

[0104] The similarity between the target node and k neighbor nodes is normalized, and smoothing is performed on the basis of the normalization process to obtain the influence weights of the k neighbor nodes on the target node.

[0105] Here, the smoothing process may be, for example, adding denominator smoothing process to the above formula (4).

[0106] For example, the above formula (4) can be further improved to obtain the following formula (5), and then the influence weights of k neighbor nodes on node u can be calculated using formula (5).

[0107] Among them, β can be a hyperparameter that can be adjusted during model training.

[0108] In addition, in some embodiments, extracting difference information from the feature information corresponding to the s neighboring nodes may specifically include:

[0109] According to the similarity between the target node and s neighbor nodes, determine the influence weight of each neighbor node on the target node among the s neighbor nodes;

[0110] Based on the influence weight of each neighbor node on the target node, the feature information corresponding to the s neighbor nodes is weighted to obtain difference information.

[0111] Based on this, in some embodiments, determining the influence weight of each of the s neighbor nodes on the target node based on the similarity between the target node and the s neighbor nodes may specifically include:

[0112] The similarity between the target node and s neighbor nodes is normalized to obtain the influence weights of the s neighbor nodes on the target node.

[0113] In addition, in order to further improve the accuracy of difference information extraction and thus improve the recommendation effect of the recommendation system, smoothing can be performed during normalization. Based on this, in some embodiments, the above-mentioned normalization of the similarity between the target node and the s neighboring nodes to obtain the influence weights of the s neighboring nodes on the target node can specifically include:

[0114] The similarity between the target node and s neighbor nodes is normalized, and smoothing is performed on the basis of the normalization process to obtain the influence weights of the s neighbor nodes on the target node.

[0115] It should be noted that the difference information The calculation method can also be the same as the above common information The calculation method is the same as that of When , the maximum similarity node set in the above formula (3), formula (4) and formula (5) is Replace with the minimum similarity node set That’s it, I won’t go into details here.

[0116] In addition, in some embodiments, determining the target feature information corresponding to the target node based on the common information and the difference information may specifically include:

[0117] A second aggregation process is performed on the common information and the difference information to obtain target feature information corresponding to the target node.

[0118] Here, the second aggregation process may be a process of aggregating the common information and the difference information into one feature information. In some embodiments, the second aggregation process may include, for example, vector feature joint processing, weighted average processing, or linear transformation processing.

[0119] For example, taking the target graph neural network layer as the nth graph neural network layer and the target node as node u, the common information and difference information of the neighbor nodes corresponding to node u can be aggregated according to the following formula (6).

[0120] in, is the feature information obtained by aggregating the common information and difference information of the neighboring nodes corresponding to node u in the nth graph neural network layer (the subscript c represents aggregation); agg() is the aggregation function, which can be, for example, a function of vector feature union, weighted average, or linear transformation; is the common information of the neighboring nodes corresponding to node u in the nth graph neural network layer; is the difference information of the neighbor nodes corresponding to node u in the nth graph neural network layer.

[0121] The above process can extract useful information from neighbor nodes and filter out neighbor nodes with smaller weights, thereby reducing the impact of over-smoothing problems caused by excessive use of feature information of neighbor nodes.

[0122] In addition, based on the above various embodiments, in some possible embodiments, the object representation model may further include a first neural network layer. The first neural network layer may be a single-layer neural network. The first neural network layer may be a neural network layer for performing dimensionality reduction processing on feature information.

[0123] On this basis, the above S150 may specifically include:

[0124] The initial node feature matrix and the feature information corresponding to the same node in the node feature matrices corresponding to multiple graph neural network layers are concatenated to obtain an aggregated feature matrix;

[0125] The first neural network layer is used to reduce the dimension of the feature information corresponding to each node in the aggregate feature matrix to obtain the target feature matrix;

[0126] The feature information corresponding to each node in the target feature matrix is ​​used as the object representation information of the object corresponding to each node, so as to obtain the object representation information corresponding to the plurality of objects respectively.

[0127] Here, the first aggregation process in the aforementioned S150 can be, for example, a splicing process. For example, if there are m graph neural network layers in the object representation model, m+1 node feature matrices (initial node feature matrix + node feature matrix output by m graph neural network layers) can be obtained. From the perspective of expanding the feature dimension corresponding to each node, the m+1 node feature matrices can be aggregated and spliced ​​to obtain an aggregate feature matrix E of (N*(n+1)d) cat , that is, the aggregate feature matrix E is obtained by splicing through the following formula (7) cat . E cat =cat(E 0 ,E 1 ,…,E m ) (7)

[0128] For example, the first neural network layer can be used to aggregate the feature matrix E cat Perform dimensionality reduction processing and map it to the final N*h target feature matrix E, where h is the feature dimension after transformation. The value of h can also be adjusted and determined during the model training process.

[0129] In addition, in some embodiments, after using the first neural network layer to perform dimensionality reduction processing on the feature information corresponding to each node in the aggregated feature matrix to obtain the target feature matrix, the object characterization method provided in the embodiment of the present application may further include:

[0130] Based on the feature information corresponding to each node in the target feature matrix, the similarity between each node and its neighboring nodes, as well as the similarity between each node and non-neighboring nodes are calculated;

[0131] Calculate the loss function value based on the similarity between each node and its neighboring nodes, as well as the similarity between each node and non-neighboring nodes;

[0132] The model parameters of the object representation model are adjusted based on the loss function value, and the topology structure and the initial node feature matrix are input into the object representation model until the object representation model converges to obtain the trained object representation model.

[0133] Based on this, the feature information corresponding to each node in the target feature matrix is ​​used as the object representation information of the object corresponding to each node to obtain the object representation information corresponding to multiple objects, which may specifically include:

[0134] The feature information corresponding to each node in the target feature matrix output by the trained object representation model is determined as the object representation information of the object corresponding to each node.

[0135] Here, during the model training process, the Bayesian loss function can be used to calculate the loss function value.

[0136] For example, the loss function may be calculated according to the following formula (8).

[0137] Among them, s ui and s uj is the similarity between nodes; (u,i) is any node u and any node i connected to it; (u,j) is any node u and any node j not connected to it, which can be obtained by random sampling, boundary sampling, etc.; λ is the regularization weight, whose value can be adjusted during the model training process; ||E|| 2 is the second-order norm of the model weight, which can be a model parameter in the object representation model; σ() is the sigmoid function.

[0138] For example, after calculating the loss function value, the model parameters in the object representation model can be adjusted based on the loss function value. In some embodiments, the model parameters can specifically include the initial node feature matrix and the network parameters corresponding to each neural network layer in the object representation model, such as the hyperparameter β involved in the smoothing process, the feature dimension h corresponding to the target feature matrix, and the regularization weight λ in the loss function.

[0139] After multiple trainings, when the object representation model converges, the training can be ended and the target feature matrix output by the trained object representation model can be obtained. The feature information corresponding to each node in the target feature matrix is ​​then determined as the object representation information of the object corresponding to each node.

[0140] In order to better describe the entire solution, some specific examples are given based on the above embodiments.

[0141] For example, a structural diagram of an object representation model is shown in Figure 2. The object representation model may include m graph neural network layers and a first neural network layer.

[0142] Based on the above object representation model, the initial node feature matrix E 0 The topological structure G is input into the first graph neural network layer, and the common information and difference information of the neighbor nodes corresponding to each node can be extracted through the first graph neural network layer. Taking node u as an example, when extracting common information, the feature information corresponding to the k neighbor nodes with the highest similarity to node u can be selected based on the attention mechanism, and the influence weight of each neighbor node among the k neighbor nodes on node u can be calculated based on the softmax function. At the same time, the denominator smoothing process is added when calculating the influence weight, and then the feature information corresponding to the k neighbor nodes is weighted summed based on the influence weight, so as to obtain the common information corresponding to the node u. Similarly, when extracting difference information, the feature information corresponding to the s neighbor nodes with the lowest similarity to node u can be selected based on the attention mechanism, and processed in the same way as above to obtain the difference information corresponding to the node u. After aggregating the common information and difference information, the feature information corresponding to the node u can be obtained. Combined with the feature information corresponding to other nodes, the node feature matrix E corresponding to the first graph neural network layer can be obtained. 1 .

[0143] Based on this, the first graph neural network layer can transform the node feature matrix E 1 Output to the vertical normalization processing module to normalize the node feature matrix E 1 Perform vertical standardization to obtain the standard node feature matrix And input it to the second layer of graph neural network layer, and then the second layer of graph neural network layer performs the above processing of extracting common information and difference information, and so on, until the mth layer of graph neural network layer outputs the node feature matrix E m .

[0144] The initial node feature matrix E 0 And the node feature matrix E output by each graph neural network layer 1 , E 2 …E mAfter aggregation, it can be input into the first neural network layer for dimensionality reduction processing to obtain the final target feature matrix E.

[0145] Based on this, on the one hand, due to the excessive reliance on and utilization of neighbor node information in existing collaborative filtering solutions based on neural networks, the "unique" information contained in each node is weakened, making the information contained in the nodes tend to be consistent, which is not conducive to recommending similar products to users. To this end, the embodiment of the present application uses an attention mechanism to extract the common information and difference information between nodes. The information after the two types of information are aggregated retains enough "unique" information, avoiding the problem of concentrated similarity distribution between nodes, thereby improving the problem of over-smoothing and improving the accuracy of object representation.

[0146] On the other hand, since the existing collaborative filtering scheme based on neural networks does not pay too much attention to the information transmission between adjacent neural network layers, and the graph neural network in each layer will narrow the similarity distribution between nodes, this will make the over-smoothing problem more serious as the number of layers increases (that is, the over-smoothing problem of the nth graph neural network layer is more serious than that of the n-1th layer). For this reason, in the embodiment of the present application, a vertical normalization process is added between adjacent graph neural network layers. Specifically, before the node feature matrix output by each graph neural network layer is passed to the next layer, the vertical normalization can correct the similarity between the nodes and distribute them in a reasonable range, so that the over-smoothing problem will not be passed to the next layer of the network, thereby further improving the over-smoothing problem and improving the accuracy of object representation.

[0147] It should be noted that the application scenarios described in the above-mentioned embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Ordinary technicians in this field can know that with the emergence of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0148] Based on the same inventive concept, the present application also provides an object representation device, which will be described in detail with reference to FIG3 .

[0149] FIG3 is a schematic structural diagram of an object representation device provided by an embodiment of the present application.

[0150] As shown in FIG3 , the object representation device 300 may include:

[0151] Relationship acquisition module 301, used to acquire association relationships between multiple objects;

[0152] A topology construction module 302 is configured to construct a topology structure with the plurality of objects as nodes according to the association relationship, wherein nodes corresponding to the objects having the association relationship in the topology structure have connecting edges;

[0153] An information input module 303 is configured to input the topological structure and the initial node feature matrix into an object representation model, wherein the object representation model includes multiple graph neural network layers, and the initial node feature matrix includes initial feature information corresponding to each node in the topological structure;

[0154] An information extraction module 304 is configured to utilize each of the multiple graph neural network layers to extract, for each node in the topological structure, common information and difference information of its neighboring nodes as feature information of the node, and output a node feature matrix corresponding to each of the multiple graph neural network layers.

[0155] The representation generation module 305 is used to perform a first aggregation process on the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers, and generate object representation information corresponding to the multiple objects.

[0156] The object representation device 300 is described in detail below.

[0157] In some embodiments, the information extraction module 304 includes:

[0158] A first determination submodule is configured to determine a target node feature matrix to be input to a target graph neural network layer based on the initial node feature matrix, where the target graph neural network layer is any one of the multiple graph neural network layers;

[0159] A first input submodule, configured to input the topological structure and the target node feature matrix into the target graph neural network layer;

[0160] The first extraction submodule is used to use the target graph neural network layer to extract the common information and difference information corresponding to the neighbor nodes of each node in the topological structure as the feature information of the node, and output a node feature matrix corresponding to the target graph neural network layer.

[0161] In some embodiments, the first determining submodule includes:

[0162] A first determining unit is configured to, when the target graph neural network layer is the first graph neural network layer among the multiple graph neural network layers, use the initial node feature matrix as a target node feature matrix to be input to the target graph neural network layer;

[0163] The second determination unit is used to use the node feature matrix corresponding to the previous graph neural network layer of the target graph neural network layer as the target node feature matrix to be input into the target graph neural network layer when the target graph neural network layer is any graph neural network layer among the multiple graph neural network layers except the first graph neural network layer.

[0164] In some embodiments, the second determining unit is specifically configured to:

[0165] For each feature dimension in the node feature matrix corresponding to the previous graph neural network layer, normalize the node feature matrix to obtain a standard node feature matrix corresponding to the previous graph neural network layer;

[0166] The standard node feature matrix is ​​used as the target node feature matrix to be input into the target graph neural network layer.

[0167] In some embodiments, the first extraction submodule includes:

[0168] A third determining unit is configured to determine, using the target graph neural network layer, t neighbor nodes corresponding to a target node from the topological structure, where the target node is any node in the topological structure;

[0169] A first acquiring unit is configured to acquire feature information corresponding to the target node and feature information corresponding to each of the t neighboring nodes from the target node feature matrix;

[0170] a first calculating unit, configured to calculate the similarity between the target node and the t neighboring nodes based on the feature information corresponding to the target node and the feature information corresponding to the t neighboring nodes;

[0171] A second acquiring unit is configured to acquire, from the t neighbor nodes, k neighbor nodes having the highest similarity to the target node and s neighbor nodes having the lowest similarity to the target node;

[0172] a first extraction unit, configured to extract the common information from the feature information corresponding to the k neighboring nodes, and extract the difference information from the feature information corresponding to the s neighboring nodes;

[0173] a fourth determining unit, configured to determine target feature information corresponding to the target node based on the commonality information and the difference information, wherein the node feature matrix corresponding to the target graph neural network layer includes the target feature information corresponding to the target node;

[0174] Wherein, t, k, and s are all integers greater than 1, and t≥k+s.

[0175] In some embodiments, the first extraction unit includes:

[0176] A first determining subunit is configured to determine an influence weight of each of the k neighboring nodes on the target node based on a similarity between the target node and the k neighboring nodes;

[0177] The first processing subunit is configured to perform weighted processing on the feature information corresponding to the k neighbor nodes based on the influence weight of each of the k neighbor nodes on the target node to obtain the common information.

[0178] In some embodiments, the first determining subunit is specifically configured to:

[0179] The similarities between the target node and the k neighbor nodes are normalized to obtain influence weights of the k neighbor nodes on the target node.

[0180] In some embodiments, the first determining subunit is further configured to:

[0181] Normalization processing is performed on the similarities between the target node and the k neighbor nodes, and smoothing processing is performed on the basis of the normalization processing to obtain the influence weights of the k neighbor nodes on the target node respectively.

[0182] In some embodiments, the fourth determining unit is specifically configured to:

[0183] A second aggregation process is performed on the common information and the difference information to obtain target feature information corresponding to the target node.

[0184] In some embodiments, the second aggregation process includes vector feature joint processing, weighted average processing or linear transformation processing.

[0185] In some embodiments, the object representation model further includes a first neural network layer;

[0186] The representation generation module 305 includes:

[0187] A first processing submodule is configured to perform information splicing processing on the feature information corresponding to the same node in the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers to obtain an aggregated feature matrix;

[0188] A second processing submodule is configured to use the first neural network layer to perform dimensionality reduction processing on the feature information corresponding to each node in the aggregated feature matrix to obtain a target feature matrix;

[0189] The second determining submodule is configured to use the feature information corresponding to each node in the target feature matrix as object representation information of the object corresponding to each node, and obtain object representation information corresponding to each of the multiple objects.

[0190] In some embodiments, the object representation device 300 further includes:

[0191] a first computing module, configured to, after performing dimensionality reduction processing on feature information corresponding to each node in the aggregated feature matrix using the first neural network layer to obtain a target feature matrix, calculate, based on the feature information corresponding to each node in the target feature matrix, the similarity between each node and its neighboring nodes, and the similarity between each node and non-neighboring nodes;

[0192] A second calculation module is used to calculate the loss function value according to the similarity between each node and its neighboring nodes, and the similarity between each node and non-neighboring nodes;

[0193] a parameter adjustment module, configured to adjust the model parameters of the object representation model based on the loss function value, and return to execute inputting the topological structure and the initial node feature matrix into the object representation model until the object representation model converges, thereby obtaining the trained object representation model;

[0194] The second determining submodule is specifically configured to:

[0195] The feature information corresponding to each node in the target feature matrix outputted by the trained object representation model is determined as the object representation information of the object corresponding to each node.

[0196] In some embodiments, the model parameters include the initial node feature matrix and network parameters corresponding to each neural network layer in the object representation model.

[0197] Thus, by utilizing the association relationship between multiple objects to construct a topological structure with objects as nodes and association relationships as connecting edges, and utilizing each graph neural network layer in the object representation model, the common information and difference information of the neighboring nodes of each node in the topological structure are extracted as the feature information of the node, and then the node feature matrix output by the object representation model corresponding to the multiple graph neural network layers contained therein is subjected to a first aggregation process, and finally the object representation information corresponding to the multiple objects is generated. In this way, since the common information and difference information of the neighboring nodes corresponding to each node are extracted in the embodiment of the present application, not only the common information between the nodes but also the difference information between different nodes is retained, the over-smoothing problem during object representation can be effectively improved, thereby improving the accuracy of object representation.

[0198] FIG4 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.

[0199] The electronic device 400 may include a processor 401 and a memory 402 storing computer program instructions.

[0200] Specifically, the processor 401 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0201] Memory 402 may include a large capacity memory for data or instructions. By way of example and not limitation, memory 402 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 402 may include removable or non-removable (or fixed) media. Where appropriate, memory 402 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, memory 402 is a non-volatile solid-state memory.

[0202] In certain embodiments, the memory may include read-only memory (ROM), random access memory (RAM), magnetic disk storage media devices, optical storage media devices, flash memory devices, electrical, optical, or other physical / tangible memory storage devices. Thus, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to an aspect of the present application.

[0203] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any one of the object representation methods in the above embodiments.

[0204] In some examples, the electronic device 400 may further include a communication interface 403 and a bus 410. As shown in FIG4, the processor 401, the memory 402, and the communication interface 403 are connected via the bus 410 and communicate with each other.

[0205] The communication interface 403 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0206] Bus 410 includes hardware, software or both, and the components of online data flow metering equipment are coupled to each other. For example, but not limitation, bus 410 may include accelerated graphics port (AGP) or other graphics bus, enhanced industry standard architecture (EISA) bus, front side bus (FSB), hypertransport (HT) interconnection, industry standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnect (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 410 may include one or more buses. Although the present application embodiment describes and shows a specific bus, the application considers any suitable bus or interconnection.

[0207] Illustratively, the electronic device 400 may be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA).

[0208] The electronic device 400 can execute the object characterization method in the embodiment of the present application, thereby realizing the object characterization method and apparatus described in conjunction with FIG. 1 and FIG. 3 .

[0209] In addition, in combination with the object characterization method in the above embodiment, the embodiment of the present application may provide a computer-readable storage medium for implementation. The computer-readable storage medium stores computer program instructions; when the computer program instructions are executed by the processor, any one of the object characterization methods in the above embodiment is implemented. Examples of computer-readable storage media include non-transitory computer-readable storage media, such as portable disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, etc.

[0210] It should be understood that the present application is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present application is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present application.

[0211] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of the present application are programs or code segments that are used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable medium" can include any medium that can store or transmit information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROMs, flash memories, erasable ROMs (EROMs), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. The code segment can be downloaded via a computer network such as the Internet, an intranet, etc.

[0212] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps. In other words, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0213] Aspects of the present application have been described above with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device to produce a machine so that these instructions executed via the processor of the computer or other programmable data processing device enable the implementation of the function / action specified in one or more boxes of the flowchart and / or block diagram. This processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor or a field programmable logic circuit. It is also understood that each box in the block diagram and / or the flowchart and the combination of the boxes in the block diagram and / or the flowchart can also be implemented by the dedicated hardware that performs the specified function or action, or can be implemented by the combination of dedicated hardware and computer instructions.

[0214] The above description is only a specific embodiment of the present application. Those skilled in the art will clearly understand that for the convenience and brevity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be included in the scope of protection of the present application.

Claims

1. An object characterization method, comprising: Get the relationship between multiple objects; Constructing a topological structure with the multiple objects as nodes according to the association relationship, wherein nodes corresponding to the objects with the association relationship in the topological structure have connecting edges; Inputting the topological structure and the initial node feature matrix into an object representation model, wherein the object representation model includes a plurality of graph neural network layers, and the initial node feature matrix includes initial feature information corresponding to each node in the topological structure; Utilizing each of the multiple graph neural network layers, for each node in the topological structure, extracting common information and difference information of its neighboring nodes as feature information of the node, and outputting node feature matrices corresponding to the multiple graph neural network layers respectively; A first aggregation process is performed on the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers to generate object representation information corresponding to the multiple objects.

2. The method according to claim 1, wherein: The method utilizes each of the multiple graph neural network layers to extract common information and difference information of neighboring nodes of each node in the topological structure as feature information of the node, and outputs node feature matrices corresponding to the multiple graph neural network layers, including: Based on the initial node feature matrix, determining a target node feature matrix to be input into a target graph neural network layer, wherein the target graph neural network layer is any one of the multiple graph neural network layers; Inputting the topological structure and the target node feature matrix into the target graph neural network layer; The target graph neural network layer is used to extract common information and difference information corresponding to neighbor nodes of each node in the topological structure as feature information of the node, and a node feature matrix corresponding to the target graph neural network layer is output.

3. The method according to claim 2, wherein: The step of determining a target node feature matrix to be input into a target graph neural network layer based on the initial node feature matrix includes: When the target graph neural network layer is the first graph neural network layer among the multiple graph neural network layers, using the initial node feature matrix as the target node feature matrix to be input into the target graph neural network layer; In the case that the target graph neural network layer is any graph neural network layer among the multiple graph neural network layers except the first graph neural network layer, the node feature matrix corresponding to the previous graph neural network layer of the target graph neural network layer is used as the target node feature matrix to be input into the target graph neural network layer.

4. The method according to claim 3, wherein: The step of using the node feature matrix corresponding to the previous graph neural network layer of the target graph neural network layer as the target node feature matrix to be input into the target graph neural network layer includes: For each feature dimension in the node feature matrix corresponding to the previous graph neural network layer, the node feature matrix is ​​standardized to obtain a standard node feature matrix corresponding to the previous graph neural network layer; The standard node feature matrix is ​​used as the target node feature matrix to be input into the target graph neural network layer.

5. The method according to claim 2, wherein: The method of using the target graph neural network layer to extract commonality information and difference information corresponding to neighboring nodes of each node in the topological structure as feature information of the node, and outputting a node feature matrix corresponding to the target graph neural network layer, includes: Using the target graph neural network layer, determining t neighbor nodes corresponding to a target node from the topological structure, wherein the target node is any node in the topological structure; Acquire feature information corresponding to the target node and feature information corresponding to the t neighboring nodes from the target node feature matrix; Calculating the similarity between the target node and the t neighbor nodes according to the feature information corresponding to the target node and the feature information corresponding to the t neighbor nodes respectively; Obtaining, from the t neighbor nodes, the k neighbor nodes with the highest similarity to the target node and the s neighbor nodes with the lowest similarity to the target node; Extracting the common information from the feature information respectively corresponding to the k neighboring nodes, and extracting the difference information from the feature information respectively corresponding to the s neighboring nodes; Determine target feature information corresponding to the target node based on the commonality information and the difference information, wherein the node feature matrix corresponding to the target graph neural network layer includes the target feature information corresponding to the target node; Wherein, t, k, and s are all integers greater than 1, and t≥k+s.

6. The method according to claim 5, wherein: The extracting the common information from the feature information respectively corresponding to the k neighboring nodes includes: Determine, according to the similarity between the target node and the k neighbor nodes, the influence weight of each neighbor node in the k neighbor nodes on the target node; Based on the influence weight of each neighbor node among the k neighbor nodes on the target node, weighted processing is performed on the feature information corresponding to the k neighbor nodes to obtain the common information.

7. The method according to claim 6, wherein: The determining, according to the similarity between the target node and the k neighbor nodes, the influence weight of each of the k neighbor nodes on the target node comprises: The similarities between the target node and the k neighbor nodes are normalized to obtain influence weights of the k neighbor nodes on the target node.

8. The method according to claim 7, wherein: The normalizing the similarity between the target node and the k neighbor nodes to obtain the influence weights of the k neighbor nodes on the target node respectively includes: The similarities between the target node and the k neighbor nodes are normalized, and smoothed based on the normalization to obtain influence weights of the k neighbor nodes on the target node.

9. The method according to claim 5, wherein: The determining the target feature information corresponding to the target node based on the common information and the difference information includes: A second aggregation process is performed on the common information and the difference information to obtain target feature information corresponding to the target node.

10. The method according to claim 9, wherein: The second aggregation processing includes vector feature joint processing, weighted average processing or linear transformation processing.

11. The method according to any one of claims 1 to 10, wherein: The object representation model also includes a first neural network layer; The first aggregation processing is performed on the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers to generate object representation information corresponding to the multiple objects, including: Performing information splicing processing on the feature information corresponding to the same node in the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers, to obtain an aggregated feature matrix; Using the first neural network layer to perform dimensionality reduction processing on feature information corresponding to each node in the aggregate feature matrix to obtain a target feature matrix; The feature information corresponding to each node in the target feature matrix is ​​used as the object representation information of the object corresponding to each node, so as to obtain the object representation information corresponding to the multiple objects respectively.

12. The method according to claim 11, wherein: After performing dimensionality reduction processing on feature information corresponding to each node in the aggregate feature matrix using the first neural network layer to obtain a target feature matrix, the method further includes: Calculate the similarity between each node and its neighbor nodes, and the similarity between each node and non-neighbor nodes according to the feature information corresponding to each node in the target feature matrix; Calculating a loss function value based on the similarity between each node and its neighboring nodes, and the similarity between each node and non-neighboring nodes; Adjusting the model parameters of the object representation model based on the loss function value, and returning to execute inputting the topological structure and the initial node feature matrix into the object representation model until the object representation model converges, thereby obtaining the trained object representation model; The step of using the feature information corresponding to each node in the target feature matrix as the object representation information of the object corresponding to each node to obtain the object representation information corresponding to the plurality of objects respectively includes: The feature information corresponding to each node in the target feature matrix outputted by the trained object representation model is determined as the object representation information of the object corresponding to each node.

13. The method according to claim 12, wherein: The model parameters include the initial node feature matrix and the network parameters corresponding to each neural network layer in the object representation model.

14. An object characterization device, comprising: A relationship acquisition module is used to obtain the association relationship between multiple objects; A topology construction module, used to construct a topology structure with the multiple objects as nodes according to the association relationship, wherein the nodes corresponding to the objects with the association relationship in the topology structure have connecting edges; An information input module, used to input the topological structure and the initial node feature matrix into an object representation model, wherein the object representation model includes a plurality of graph neural network layers, and the initial node feature matrix includes initial feature information corresponding to each node in the topological structure; An information extraction module, used to use each of the multiple graph neural network layers to extract common information and difference information of neighboring nodes of each node in the topological structure as feature information of the node, and output node feature matrices corresponding to the multiple graph neural network layers respectively; The representation generation module is used to perform a first aggregation process on the initial node feature matrix and the node feature matrices corresponding to the multiple graph neural network layers, so as to generate object representation information corresponding to the multiple objects.

15. An electronic device, comprising: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, the steps of the object representation method according to any one of claims 1 to 13 are implemented.

16. A computer-readable storage medium, wherein computer program instructions are stored on the computer-readable storage medium, and when the computer program instructions are executed by a processor, the steps of the object representation method according to any one of claims 1 to 13 are implemented.

Images