A user relationship mining method and system based on human network

Abstract

The present invention discloses a user relationship mining method and system based on human network, and proposes a relationship feature extraction network. By integrating the weight distribution characteristics of the attention mechanism and the graph structure learning ability of GCN, the accuracy of node feature extraction and relationship prediction in the user relationship network is effectively improved. First, the graph structure learning and feature deep mining are carried out in the GCN model, and the importance of nodes in the user relationship network and the attribute information and interaction information in the nodes are evaluated by combining the attention mechanism. At the same time, the mutual influence relationship between nodes is mined by combining the GCN model and the attention mechanism. The weight of the GCN model is adjusted again based on the mutual influence relationship, and finally the accurate mining and prediction of user relationships are realized. It can effectively mine the hidden complex relationships between users. The model has strong flexibility and can be adjusted and expanded according to different application scenarios. It is suitable for different types of user data analysis tasks.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a method and system for mining user relationships based on human networking. Background Art

[0002] In the context of the digital age, connections between people have transcended traditional boundaries, forming an intricate network of people. This network is not limited to physical contact, but is also expanded and deepened through virtual social platforms, mobile communication networks, and e-commerce activities. However, current user relationship mining technologies mostly focus on the analysis of a single data source, such as social network interaction records or purchase behavior data. This limitation leads to a one-sided and superficial understanding of user relationships. Therefore, there is an urgent need for a comprehensive method that can transcend data silos and deeply analyze user relationships to reveal the deep connections and potential value hidden behind diverse interactions. Summary of the invention

[0003] The purpose of the present invention is to provide a user relationship mining method and system based on human networking to solve the above-mentioned problems existing in the prior art.

[0004] An embodiment of the present invention provides a method for mining user relationships based on a human network, the method comprising:

[0005] Construct a user interaction graph: With users as user nodes and interactions between users as edges, construct a user interaction graph containing all user nodes. In the user interaction graph, the attribute vector of each user node includes the attribute information of the user and the statistical information of the user's interaction with other users. The attribute vector of the edge includes the interaction activity information between two users.

[0006] User relationship mining: Through the relationship feature extraction network, based on the user interaction graph, obtain the relationship feature graph; based on the relationship feature graph, obtain the user relationship graph; the user node in the user relationship graph represents the user, and the edge weight in the relationship graph represents the relationship type between the two users connected by the edge; the relationship types include relatives, friends, colleagues, and customers;

[0007] The relationship feature extraction network includes a first network, a second network, a third network and an attention adjustment structure; the first network is used to extract features from attribute vectors of user nodes in a user interaction graph, and the second network is used to extract features from attribute vectors of edges in the user interaction graph; the third network is used to extract features from the user interaction graph, and the attention adjustment structure is used to generate an attention adjustment matrix, which is used to adjust the weight matrix of the third network.

[0008] Optionally, a relational feature graph is obtained based on the user interaction graph through a relational feature extraction network, including:

[0009] Input the user interaction graph into the relationship feature extraction network, perform feature extraction on the attribute vector of the user node in the user interaction graph through the first network to obtain an attribute feature vector, the attribute feature vector includes attribute features; perform feature extraction on the attribute vector of the edge in the user interaction graph through the second network to obtain an edge feature vector, the edge feature vector includes edge features; perform feature extraction on the user interaction graph through the third network to obtain an overall relationship feature graph; the overall relationship feature graph includes feature nodes and feature edges connecting two feature nodes;

[0010] For a user node in a user interaction graph, a first attention vector of the user node is constructed, where the first attention vector includes a first attention index; the first attention index represents the degree of attention of the user node to the adjacent user nodes of the user node, where the adjacent user nodes of the user node are other user nodes connected to the user node;

[0011] For a feature node in the overall relationship feature graph, construct a second attention vector of the feature node; the second attention vector includes a second attention index, and the second attention index represents the degree of attention of the feature node to the adjacent feature nodes of the feature node, and the adjacent feature nodes of the feature node are other feature nodes connected to the feature node;

[0012] Based on the overall relationship feature graph and the user interaction graph, a third attention vector of the user node is obtained; the third attention vector includes a third attention index; the third attention index indicates the degree of influence of the user node on the feature node corresponding to the user node;

[0013] Obtain the loss function of the third network;

[0014] If the loss function is greater than a first set value, and the first attention vector and the second attention vector satisfy a first condition, constructing an attention adjustment matrix based on the first attention vector, the second attention vector, and the third attention vector;

[0015] Adjusting a weight matrix of a third network based on the attention adjustment matrix;

[0016] If the loss function is less than or equal to the first set value, the overall relationship feature graph output by the third network is used as the relationship feature graph; the relationship feature graph includes interconnected nodes, node attribute feature information, and relationship feature information of edges between interconnected nodes.

[0017] Optionally, based on the overall relationship feature graph and the user interaction graph, a third attention vector of the user node is obtained; the third attention vector includes a third attention index; the third attention index represents the degree of influence of the user node on the feature node corresponding to the user node, including:

[0018] For user node k, obtain the feature node corresponding to user node k in the overall relationship feature graph as the target node, and the target node and the user node correspond to the same user; obtain other feature nodes connected to the target node in the overall relationship feature graph; use the target node and other feature nodes connected to the target node as matching feature nodes of user node k;

[0019] Obtain the information vector of user node k and the information vectors of all user nodes connected to user node k, where the information is obtained by concatenating the attribute vector of the user node and the attribute vector of the edge; construct a node adjacent matrix based on the information vectors of user node k and all user nodes connected to user node k, where each information vector constitutes a row of the node adjacent matrix;

[0020] For each matching feature node x, obtain a feature information vector of the matching feature node x, the feature information vector being composed of the node attribute feature information and the relationship feature information; and construct a feature adjacent matrix by taking all feature information vectors as rows of the feature adjacent matrix;

[0021] The exchange entropy between the feature adjacency matrix and the node adjacency matrix is ​​used as the third attention index.

[0022] Optionally, based on the relationship feature graph, a user relationship graph is obtained, including:

[0023] Based on the node attribute feature information and the relationship feature information, the relationship type between the interconnected nodes is obtained;

[0024] The relationship types are marked on the edges of the relationship feature graph to obtain the user relationship graph.

[0025] Optionally, if the loss function is greater than a first set value, and the first attention vector and the second attention vector satisfy a first condition, constructing an attention adjustment matrix based on the first attention vector, the second attention vector, and the third attention vector includes:

[0026] Obtaining matching point pairs, the matching point pairs including user nodes in the user interaction graph and feature nodes in the overall relationship feature graph that match each other;

[0027] For each matching point pair, obtain the first cross entropy between the first attention vector of the user node and the second attention vector of the feature node; obtain multiple first cross entropies corresponding to the user interaction graph; obtain the average of the multiple first cross entropies;

[0028] If the loss function is greater than the first set value, and the average of the multiple first cross entropies is greater than the second set value, for each user node, the first attention vector, the second attention vector, and the third attention vector corresponding to the user node are mapped to the same space to obtain a first mapping vector, a second mapping vector, and a third mapping vector, respectively;

[0029] Performing a weighted summation on the first mapping vector, the second mapping vector, and the third mapping vector to obtain an attention adjustment vector of the user node; the attention adjustment vector includes a plurality of attention adjustment factors, each attention adjustment factor corresponding to an edge connected to the user node;

[0030] Multiple attention adjustment vectors are arranged in sequence to form an attention adjustment matrix.

[0031] Optionally, the user interaction graph is constructed in the form of a weighted adjacency list.

[0032] The embodiment of the present invention further provides a user relationship mining system based on human networking, the system comprising:

[0033] The user interaction graph construction module is used to construct a user interaction graph containing all user nodes, with users as user nodes and interactions between users as edges. In the user interaction graph, the attribute vector of each user node includes the attribute information of the user and the statistical information of the user's interaction with other users; the attribute vector of the edge includes the interaction activity information between two users.

[0034] The user relationship mining module is used to extract the network through the relationship feature, obtain the relationship feature graph based on the user interaction graph, and obtain the user relationship graph based on the relationship feature graph; the user node in the user relationship graph represents the user, and the weight of the edge in the relationship graph represents the relationship type between the two users connected by the edge; the relationship type includes relatives, friends, colleagues, and customers;

[0035] The relationship feature extraction network includes a first network, a second network, a third network and an attention adjustment structure; the first network is used to extract features from attribute vectors of user nodes in a user interaction graph, and the second network is used to extract features from attribute vectors of edges in the user interaction graph; the third network is used to extract features from the user interaction graph, and the attention adjustment structure is used to generate an attention adjustment matrix, which is used to adjust the weight matrix of the third network.

[0036] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0037] The embodiment of the present invention provides a user relationship mining method and system based on the human network, and proposes a relationship feature extraction network, which includes a first network, a second network, a third network and an attention adjustment structure. Among them, the first network is used to extract features from the attribute vectors of the user nodes in the user interaction graph, the second network is used to extract features from the attribute vectors of the edges in the user interaction graph, the third network is used to extract features from the user interaction graph, the attention adjustment structure is used to generate an attention adjustment matrix, and the attention adjustment matrix is ​​used to adjust the weight matrix of the third network. In an embodiment of the present invention, the first network, the second network, and the third network are all graph convolution networks (GraphConvolution Network, GCN). It is necessary to first train the first network, the second network, and the third network separately, and then construct a relationship feature extraction network with the trained first network, the second network, and the third network. The relationship feature extraction network does not need to be trained separately when in use, and the user interaction graph is directly input into the relationship feature extraction network, and the relationship feature extraction network realizes self-adjustment of the weight matrix through the attention adjustment structure.

[0038] By adopting the above scheme, by integrating the weight distribution characteristics of the attention mechanism and the graph structure learning ability of GCN, the accuracy of node feature extraction and relationship prediction in the user relationship network is effectively improved. First, the graph structure learning and feature deep mining are carried out in the GCN model, and the importance of the nodes in the user relationship network and the attribute information and interaction information in the nodes are evaluated by combining the attention mechanism. At the same time, the mutual influence relationship between nodes is mined by combining the GCN model and the attention mechanism. The weight of the GCN model is adjusted again based on the mutual influence relationship, and finally the accurate mining and prediction of user relationships are realized. Able to mine complex relationships: The model can effectively mine the hidden complex relationships between users, thereby improving the accuracy of the recommendation system and the effect of social network analysis. In addition, this application does not require additional training of the entire relationship feature extraction network. The relationship feature extraction network can adjust the weights according to the implemented detection data, so that the model can be applied to different input data and improve the generalization ability of the model. Flexibility: The model has strong flexibility and can be adjusted and expanded according to different application scenarios, and is suitable for different types of user data analysis tasks.

[0039] In summary, the relationship feature extraction network (model) provided by this application combines the attention mechanism and the graph convolutional network (GCN), which can learn the complex relationships between users from user behavior data. This model not only considers the direct interaction between users, but also captures the high-order relationships hidden behind user behavior, thereby improving the accuracy of the recommendation system and the effect of social network analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1It is a flow chart of a method for mining user relationships based on human networking provided by an embodiment of the present invention.

[0041] Figure 2 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention.

[0042] Markings in the figure: 500 - bus; 501 - receiver; 502 - processor; 503 - transmitter; 504 - memory; 505 - bus interface. DETAILED DESCRIPTION

[0043] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0044] Example

[0045] The embodiment of the present invention provides a method for mining user relationships based on human networking. Figure 1 As shown, the method includes:

[0046] S101: Construct a user interaction graph: With users as user nodes and interactions between users as edges, a user interaction graph containing all user nodes is constructed. In the user interaction graph, the attribute vector of each user node includes the attribute information of the user and the statistical information of the user's interaction with other users; the attribute vector of the edge includes the interactive activity information between two users. The attribute information of the user includes the user identity information (which can also be the user ID), age, gender, interests and other information that characterizes the user's identity characteristics. The statistical information of the user's interaction with other users includes the number of interactions, interaction time, interaction frequency, interaction type distribution, etc. The interactive activity information between two users includes likes, comments, forwarding, purchases, transfers, chats, etc.

[0047] In an embodiment of the present invention, a user interaction graph can be constructed in the form of a weighted adjacency list, that is, by constructing a weighted adjacency list of user nodes, wherein the weight of the weighted adjacency list is the interaction activity information between two users. Optionally, the weight of the weighted adjacency list can be obtained by weighting the interaction activity information between two users.

[0048] S102: User relationship mining: extracting the relationship feature network through the relationship feature, obtaining the relationship feature graph based on the user interaction graph; and obtaining the user relationship graph based on the relationship feature graph.

[0049] Among them, the user node in the user relationship graph represents the user, and the weight of the edge in the relationship graph represents the relationship type between two users connected by the edge; the relationship types include relatives, friends, colleagues, customers, etc. Different relationship types are represented by different numerical symbols, for example, 1 represents relatives, 2 represents friends, 3 represents colleagues, and 4 represents customers.

[0050] In an embodiment of the present invention, the relationship feature extraction network includes a first network, a second network, a third network and an attention adjustment structure. Among them, the first network is used to extract features from the attribute vectors of the user nodes in the user interaction graph, the second network is used to extract features from the attribute vectors of the edges in the user interaction graph, the third network is used to extract features from the user interaction graph, and the attention adjustment structure is used to generate an attention adjustment matrix, and the attention adjustment matrix is ​​used to adjust the weight matrix of the third network. In an embodiment of the present invention, the first network, the second network, and the third network are all graph convolution networks (GraphConvolution Network, GCN). It is necessary to first train the first network, the second network, and the third network separately, and then construct a relationship feature extraction network with the trained first network, the second network, and the third network. The relationship feature extraction network does not need additional training when in use, and the user interaction graph is directly input into the relationship feature extraction network, and the relationship feature extraction network realizes self-adjustment of the weight matrix through the attention adjustment structure.

[0051] By adopting the above scheme, by integrating the weight distribution characteristics of the attention mechanism and the graph structure learning ability of GCN, the accuracy of node feature extraction and relationship prediction in the user relationship network is effectively improved. First, the graph structure learning and feature deep mining are carried out in the GCN model, and the importance of the nodes in the user relationship network and the attribute information and interaction information in the nodes are evaluated by combining the attention mechanism. At the same time, the mutual influence relationship between nodes is mined by combining the GCN model and the attention mechanism. The weight of the GCN model is adjusted again based on the mutual influence relationship, and finally the accurate mining and prediction of user relationships are realized. Able to mine complex relationships: The model can effectively mine the hidden complex relationships between users, thereby improving the accuracy of the recommendation system and the effect of social network analysis. In addition, this application does not require additional training of the entire relationship feature extraction network. The relationship feature extraction network can adjust the weights according to the implemented detection data, so that the model can be applied to different input data and improve the generalization ability of the model. Flexibility: The model has strong flexibility and can be adjusted and expanded according to different application scenarios, and is suitable for different types of user data analysis tasks.

[0052] In summary, the relationship feature extraction network (model) provided by this application combines the attention mechanism and the graph convolutional network (GCN), which can learn the complex relationships between users from user behavior data. This model not only considers the direct interaction between users, but also captures the high-order relationships hidden behind user behavior, thereby improving the accuracy of the recommendation system and the effect of social network analysis.

[0053] The training method of the first network includes:

[0054] A first training set is obtained, the first training set includes multiple training attribute vectors of multiple nodes, and features are extracted from the training set through the first network. If the loss function of the first network converges, it is determined that the training of the first network is finished. The training attribute vector includes user identity information (which may also be a user ID), age, gender, interests, etc.

[0055] The training method of the second network includes:

[0056] A second training set is obtained, the second training set includes a plurality of training attribute vectors, the training attribute vectors correspond to edges, and each edge connects two nodes. The second training set is subjected to feature extraction through the second network, and if the loss function of the second network converges, it is determined that the training of the second network is finished. The training attribute vectors include likes, comments, forwarding, purchases, transfers, chats, etc.

[0057] The training method of the third network includes:

[0058] A third training set is obtained, and the third training set includes multiple training user interaction graphs. In the training user interaction graphs, the attribute vector of each user node includes the user's training attribute information and the statistical information of the user's interaction with other users; the edge's training attribute vector includes the interactive activity information between two users. The user's attribute information includes user identity information (which may also be user ID), age, gender, interests, and other information that characterizes the user's identity characteristics. The statistical information of the user's interaction with other users includes the number of interactions, interaction time, interaction frequency, and interaction type distribution. The third training set is subjected to feature extraction through the third network. If the loss function of the third network converges, it is determined that the training of the third network is completed.

[0059] In an embodiment of the present invention, the loss functions of the first network, the second network and the third network may be a cross entropy loss function (Cross Entropy Loss), a multi-category cross entropy loss function (Categorical Cross Entropy Loss), a binary cross entropy loss function (Binary Cross Entropy Loss), or a mean square error loss function (MeanSquared Error Loss).

[0060] In an embodiment of the present invention, a relationship feature graph is obtained based on a user interaction graph through a relationship feature extraction network, including:

[0061] The user interaction graph is input into the relationship feature extraction network, and the attribute vector of the user node in the user interaction graph is subjected to feature extraction through the first network to obtain an attribute feature vector, which includes attribute features. The attribute vector of the edge in the user interaction graph is subjected to feature extraction through the second network to obtain an edge feature vector, which includes edge features. The user interaction graph is subjected to feature extraction through the third network to obtain an overall relationship feature graph, which includes feature nodes and feature edges connecting two feature nodes.

[0062] For a user node in a user interaction graph, a first attention vector of the user node is constructed, and the first attention vector includes a first attention index. The first attention index represents the degree of attention of the user node to the neighboring user nodes of the user node. The higher the degree of attention, the higher the interaction index: the more frequent the interaction frequency, the more intimate the depth of the interaction type. The neighboring user nodes of the user node are other user nodes connected to the user node.

[0063] Among them, each user node corresponds to a first attention vector, and each first attention index corresponds to an edge.

[0064] In an embodiment of the present invention, it also includes:

[0065] Through the multiple attention heads in the attention adjustment structure, each attention head corresponds to an edge, and the first attention index corresponding to an edge in each user node through the attention mechanism is obtained as follows:

[0066] For user node k, obtain the attribute vector of the mth edge of user node k, and for the i-th attribute l of the attribute vector i The attribute vector g of the nth edge of user node k, n is different from m, and the i-th l is calculated i Similarity score e with g i =f(l i , g), where f(l i , g) indicates l i The point convolution operation is performed with g; where k is a positive integer ranging from 1 to K, K is the number of user nodes, that is, the number of users, and is a positive integer. i, n, and m are positive integers ranging from 1 to N, N is the number of other user nodes connected to user node k, that is, the number of edges connected to user node k.

[0067] The first attention index is obtained based on the similarity score, specifically:

[0068] Among them, α i represents the first attention index of the i-th attribute, e j Represents the jth attribute l jSimilarity score with attribute vector g; e i Represents the i-th attribute l i The similarity score with the attribute vector g; J represents the number of local features, and the value of J is the same as N.

[0069] For a feature node in the overall relationship feature graph, a second attention vector of the feature node is constructed; the second attention vector includes a second attention index, and the second attention index represents the degree of attention of the feature node to the adjacent feature nodes of the feature node, and the adjacent feature nodes of the feature node are other feature nodes connected to the feature node.

[0070] In an embodiment of the present invention, a method for obtaining the second attention index is similar to a method for obtaining the first attention index. For details, reference may be made to the method for obtaining the first attention index, which will not be repeated here.

[0071] Based on the overall relationship feature graph and the user interaction graph, the third attention vector of the user node is obtained; the third attention vector includes a third attention index. The third attention index indicates the degree of influence of the user node on the feature node corresponding to the user node. The third attention index is obtained as follows:

[0072] For user node k, the feature node corresponding to user node k in the overall relationship feature graph is used as the target node. The target node and the user node correspond to the same user and can be considered as the same node, except that the target node is in the overall relationship feature graph and the user node k is in the user interaction graph. Other feature nodes (neighboring feature nodes) connected to the target node in the overall relationship feature graph are obtained. The target node and other feature nodes connected to the target node are used as matching feature nodes for user node k.

[0073] The information vector of user node k and the information vectors of all user nodes connected to user node k are obtained. The information is obtained by concatenating the attribute vector of the user node and the attribute vector of the edge. A node adjacent matrix is ​​constructed based on the information vectors of user node k and all user nodes connected to user node k. Each information vector constitutes a row of the node adjacent matrix.

[0074] For each matching feature node x, obtain the feature information vector of the matching feature node x, which is composed of the node attribute feature information and the relationship feature information. All feature information vectors are used as rows of the feature adjacency matrix to form the feature adjacency matrix. The overall relationship feature graph contains the interconnected nodes, node attribute feature information, and the relationship feature information of the edges between the interconnected nodes.

[0075] The exchange entropy (or cross entropy) between the feature adjacency matrix and the node adjacency matrix is ​​used as the third attention index.

[0076] If the loss function is greater than the first set value (the first set value can be between 0.1 and 0.5), and the first attention vector and the second attention vector meet the first condition, an attention adjustment matrix is ​​constructed based on the first attention vector, the second attention vector and the third attention vector. Specifically, the first attention vector, the second attention vector and the third attention vector are weighted and summed to obtain a fourth attention vector; each user node corresponds to a fourth attention vector, and according to the position of the user node in the graph, the fourth attention vector is set in the position to obtain an attention adjustment matrix.

[0077] If the loss function is greater than a first set value, and the first attention vector and the second attention vector satisfy a first condition, an attention adjustment matrix is ​​constructed based on the first attention vector, the second attention vector, and the third attention vector, including:

[0078] Obtaining matching point pairs, the matching point pairs including user nodes in the user interaction graph and feature nodes in the overall relationship feature graph that match each other;

[0079] For each matching point pair, obtain the first cross entropy between the first attention vector of the user node and the second attention vector of the feature node; obtain multiple first cross entropies corresponding to the user interaction graph; obtain the average of multiple first cross entropies; if the loss function is greater than the first set value, and the average of multiple first cross entropies is greater than the second set value (the second set value is 0.2 to 0.5),

[0080] For each user node, the first attention vector, the second attention vector, and the third attention vector corresponding to the user node are mapped into the same space to obtain a first mapping vector, a second mapping vector, and a third mapping vector respectively;

[0081] A weighted sum is performed on the first mapping vector, the second mapping vector, and the third mapping vector to obtain an attention adjustment vector of the user node; the attention adjustment vector includes a plurality of attention adjustment factors, and each attention adjustment factor corresponds to an edge connected to the user node.

[0082] Multiple attention adjustment vectors are arranged in sequence to form an attention adjustment matrix.

[0083] The weight matrix of the third network is adjusted based on the attention adjustment matrix. Specifically, according to the user nodes corresponding to the fourth attention vector obtained above, and the user nodes corresponding to the elements in the fourth attention vector, the weights of the third network corresponding to the user nodes are adjusted. Specifically, the values ​​of the elements in the fourth attention vector are used as adjustment factors to adjust the weights of the third network: new weight = value of the element in the fourth attention vector * weight. In an embodiment of the present invention, the weight matrix of the third network is adjusted based on the attention adjustment matrix through the attention mechanism.

[0084] After adjusting the weight matrix of the third network, feature extraction is performed again on the user interaction graph through the third network until the loss function of the third network is less than or equal to the first set value.

[0085] If the loss function is less than or equal to the first set value, the overall relationship feature graph output by the third network is used as the relationship feature graph; the relationship feature graph includes interconnected nodes, node attribute feature information, and relationship feature information of edges between interconnected nodes.

[0086] Optionally, based on the relationship feature graph, a user relationship graph is obtained, including:

[0087] Based on the node attribute feature information and the relationship feature information, the relationship type between the interconnected nodes is obtained. In an embodiment of the present invention, candidate relationship types matching the node attribute feature information and the relationship feature information are stored in the database in advance. If the node attribute feature information and the relationship feature information match the candidate relationship type, the candidate relationship type is used as the relationship type between the interconnected nodes. The relationship type is marked in the edge of the relationship feature graph to obtain a user relationship graph. For example, the mark 1 represents a relative, the mark 2 represents a friend, the mark 3 represents a colleague, and the mark 4 represents a customer.

[0088] By adopting the above scheme, we can repeatedly find out the implicit relationship between user nodes in the user interaction graph, the implicit relationship between feature nodes in the overall relationship feature graph, and the implicit relationship between user nodes in the user interaction graph and feature nodes in the overall relationship feature graph. Based on this, user relationship mining can better capture the complex relationships between users and improve the accuracy of relationship mining.

[0089] Based on the above-mentioned user relationship mining method based on human networking, an embodiment of the present invention provides a user relationship mining system based on human networking, which is used to execute the above-mentioned user relationship mining method based on human networking. The user relationship mining system based on human networking includes:

[0090] The user interaction graph construction module is used to construct a user interaction graph containing all user nodes, with users as user nodes and interactions between users as edges. In the user interaction graph, the attribute vector of each user node includes the attribute information of the user and the statistical information of the user's interaction with other users; the attribute vector of the edge includes the interaction activity information between two users.

[0091] The user relationship mining module is used to extract the network through the relationship feature, obtain the relationship feature graph based on the user interaction graph, and obtain the user relationship graph based on the relationship feature graph; the user node in the user relationship graph represents the user, and the weight of the edge in the relationship graph represents the relationship type between the two users connected by the edge; the relationship type includes relatives, friends, colleagues, and customers;

[0092] The relationship feature extraction network includes a first network, a second network, a third network and an attention adjustment structure; the first network is used to extract features from attribute vectors of user nodes in a user interaction graph, and the second network is used to extract features from attribute vectors of edges in the user interaction graph; the third network is used to extract features from the user interaction graph, and the attention adjustment structure is used to generate an attention adjustment matrix, which is used to adjust the weight matrix of the third network.

[0093] The specific implementation methods of the functions of each module of the above system are as described in the user relationship mining method based on the human network, and will not be repeated here.

[0094] The embodiment of the present invention also provides an electronic device for integrating the above-mentioned user relationship mining system based on human networking, such as Figure 2 As shown, the electronic device includes a memory 504, a processor 502, and a computer program stored in the memory 504 and executable on the processor 502. When the processor 502 executes the program, the steps of any of the user relationship mining methods based on human networking described above are implemented.

[0095] Among them, Figure 2In the embodiment of the present invention, a bus architecture (represented by bus 500) is shown, which may include any number of interconnected buses and bridges, and bus 500 links various circuits including one or more processors represented by processor 502 and memory represented by memory 504. Bus 500 may also link various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. Bus interface 505 provides an interface between bus 500 and receiver 501 and transmitter 503. Receiver 501 and transmitter 503 may be the same element, namely a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 502 is responsible for managing bus 500 and general processing, while memory 504 may be used to store data used by processor 502 when performing operations.

[0096] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of any method of the user relationship mining method based on human networking described above are implemented.

[0097] The algorithm and display provided herein are not inherently related to any particular computer, virtual system or other device. Various general purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious that the structure required for constructing such systems. In addition, the present invention is not directed to any specific programming language either. It should be understood that various programming languages ​​can be utilized to realize the content of the present invention described herein, and the description of the above specific languages ​​is for disclosing the best mode of the present invention.

[0098] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0099] Similarly, it should be understood that in order to streamline the present disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, various features of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting the intention that the claimed invention requires more features than those explicitly recited in each claim. More specifically, as reflected in the claims below, inventive aspects lie in less than all the features of the individual embodiments disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, with each claim itself serving as a separate embodiment of the present invention.

[0100] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and in addition they may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed in this manner may be combined in any combination. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0101] In addition, those skilled in the art will appreciate that, although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, in the claims below, any one of the claimed embodiments may be used in any combination.

[0102] The various component embodiments of the present invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components in the apparatus according to an embodiment of the present invention. The present invention may also be implemented as a device or device program (e.g., a computer program and a computer program product) for executing part or all of the methods described herein. Such a program implementing the present invention may be stored on a computer-readable medium, or may have the form of one or more signals. Such a signal may be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

Claims

1. A user relationship mining method based on human networking, characterized in that: The method comprises: Construct a user interaction graph: With users as user nodes and interactions between users as edges, construct a user interaction graph containing all user nodes. In the user interaction graph, the attribute vector of each user node includes the attribute information of the user and the statistical information of the user's interaction with other users. The attribute vector of the edge includes the interaction activity information between two users. User relationship mining: Through the relationship feature extraction network, based on the user interaction graph, obtain the relationship feature graph; based on the relationship feature graph, obtain the user relationship graph; the user node in the user relationship graph represents the user, and the edge weight in the relationship graph represents the relationship type between the two users connected by the edge; the relationship types include relatives, friends, colleagues, and customers; The relationship feature extraction network includes a first network, a second network, a third network and an attention adjustment structure; the first network is used to extract features from attribute vectors of user nodes in a user interaction graph, and the second network is used to extract features from attribute vectors of edges in the user interaction graph; the third network is used to extract features from the user interaction graph, and the attention adjustment structure is used to generate an attention adjustment matrix, and the attention adjustment matrix is ​​used to adjust the weight matrix of the third network; Through the relationship feature extraction network, based on the user interaction graph, a relationship feature graph is obtained, including: Input the user interaction graph into the relational feature extraction network to obtain an overall relational feature graph; the overall relational feature graph includes feature nodes and feature edges connecting two feature nodes; For a user node in a user interaction graph, construct a first attention vector of the user node, where the first attention vector includes a first attention index; the first attention index represents the degree of attention of the user node to the adjacent user nodes of the user node; For a feature node in the overall relationship feature graph, construct a second attention vector of the feature node; the second attention vector includes a second attention index, and the second attention index represents the degree of attention of the feature node to the adjacent feature nodes of the feature node; Based on the overall relationship feature graph and the user interaction graph, a third attention vector of the user node is obtained; the third attention vector includes a third attention index; the third attention index indicates the degree of influence of the user node on the feature node corresponding to the user node; and a loss function of the third network is obtained; If the loss function is greater than a first set value, and the first attention vector and the second attention vector satisfy a first condition, constructing an attention adjustment matrix based on the first attention vector, the second attention vector, and the third attention vector; Adjusting a weight matrix of a third network based on the attention adjustment matrix; If the loss function is less than or equal to the first set value, the overall relationship feature graph output by the third network is used as the relationship feature graph.

2. The user relationship mining method based on human networking according to claim 1 is characterized in that: Based on the overall relationship feature graph and the user interaction graph, the third attention vector of the user node is obtained; The third attention vector includes a third attention index; The third attention index indicates the degree of influence of the user node on the feature node corresponding to the user node, including: For user node k, obtain the feature node corresponding to user node k in the overall relationship feature graph as the target node, and the target node and the user node correspond to the same user; obtain other feature nodes connected to the target node in the overall relationship feature graph; use the target node and other feature nodes connected to the target node as matching feature nodes of user node k; Obtain the information vector of user node k and the information vectors of all user nodes connected to user node k, where the information is obtained by concatenating the attribute vector of the user node and the attribute vector of the edge; construct a node adjacent matrix based on the information vectors of user node k and all user nodes connected to user node k, where each information vector constitutes a row of the node adjacent matrix; For each matching feature node x, obtain a feature information vector of the matching feature node x, the feature information vector being composed of the node attribute feature information and the relationship feature information; and construct a feature adjacent matrix by taking all feature information vectors as rows of the feature adjacent matrix; The exchange entropy between the feature adjacency matrix and the node adjacency matrix is ​​used as the third attention index.

3. The user relationship mining method based on human networking according to claim 1 is characterized in that: Based on the relationship feature graph, the user relationship graph is obtained, including: Based on the node attribute feature information and the relationship feature information, the relationship type between the interconnected nodes is obtained; The relationship types are marked on the edges of the relationship feature graph to obtain the user relationship graph.

4. The user relationship mining method based on human networking according to claim 1 is characterized in that: If the loss function is greater than the first set value, and the first attention vector and the second attention vector meet the first condition, an attention adjustment matrix is ​​constructed based on the first attention vector, the second attention vector and the third attention vector, including: Obtaining matching point pairs, the matching point pairs including user nodes in the user interaction graph and feature nodes in the overall relationship feature graph that match each other; For each matching point pair, obtain the first cross entropy between the first attention vector of the user node and the second attention vector of the feature node; obtain multiple first cross entropies corresponding to the user interaction graph; obtain the average of the multiple first cross entropies; If the loss function is greater than the first set value, and the average of the multiple first cross entropies is greater than the second set value, for each user node, the first attention vector, the second attention vector, and the third attention vector corresponding to the user node are mapped to the same space to obtain a first mapping vector, a second mapping vector, and a third mapping vector, respectively; Performing a weighted summation on the first mapping vector, the second mapping vector, and the third mapping vector to obtain an attention adjustment vector of the user node; the attention adjustment vector includes a plurality of attention adjustment factors, each attention adjustment factor corresponding to an edge connected to the user node; Multiple attention adjustment vectors are arranged in sequence to form an attention adjustment matrix.

5. The user relationship mining method based on human networking according to claim 4 is characterized in that: The user interaction graph is constructed in the form of a weighted adjacency list.

6. A user relationship mining system based on human networking, characterized in that: The system comprises: The user interaction graph construction module is used to construct a user interaction graph containing all user nodes, with users as user nodes and interactions between users as edges. In the user interaction graph, the attribute vector of each user node includes the attribute information of the user and the statistical information of the user's interaction with other users; the attribute vector of the edge includes the interaction activity information between two users. A user relationship mining module is used to obtain a relationship feature graph based on a user interaction graph through a relationship feature extraction network; and to obtain a user relationship graph based on the relationship feature graph; a user node in the user relationship graph represents a user, and the weight of an edge in the relationship graph represents the type of relationship between two users connected by the edge; the relationship types include relatives, friends, colleagues, and customers; the relationship feature extraction network includes a first network, a second network, a third network, and an attention adjustment structure; the first network is used to extract features from attribute vectors of user nodes in the user interaction graph, and the second network is used to extract features from attribute vectors of edges in the user interaction graph; the third network is used to extract features from the user interaction graph, and the attention adjustment structure is used to generate an attention adjustment matrix, and the attention adjustment matrix is ​​used to adjust the weight matrix of the third network; Through the relationship feature extraction network, based on the user interaction graph, a relationship feature graph is obtained, including: Input the user interaction graph into the relational feature extraction network to obtain an overall relational feature graph; the overall relational feature graph includes feature nodes and feature edges connecting two feature nodes; For a user node in a user interaction graph, construct a first attention vector of the user node, where the first attention vector includes a first attention index; the first attention index represents the degree of attention of the user node to the adjacent user nodes of the user node; For a feature node in the overall relationship feature graph, construct a second attention vector of the feature node; the second attention vector includes a second attention index, and the second attention index represents the degree of attention of the feature node to the adjacent feature nodes of the feature node; Based on the overall relationship feature graph and the user interaction graph, a third attention vector of the user node is obtained; the third attention vector includes a third attention index; the third attention index indicates the degree of influence of the user node on the feature node corresponding to the user node; and a loss function of the third network is obtained; If the loss function is greater than a first set value, and the first attention vector and the second attention vector satisfy a first condition, constructing an attention adjustment matrix based on the first attention vector, the second attention vector, and the third attention vector; Adjusting a weight matrix of a third network based on the attention adjustment matrix; If the loss function is less than or equal to the first set value, the overall relationship feature graph output by the third network is used as the relationship feature graph.

7. The user relationship mining system based on human networking according to claim 6 is characterized in that: Based on the overall relationship feature graph and the user interaction graph, the third attention vector of the user node is obtained; The third attention vector includes a third attention index; The third attention index indicates the degree of influence of the user node on the feature node corresponding to the user node, including: For user node k, obtain the feature node corresponding to user node k in the overall relationship feature graph as the target node, and the target node and the user node correspond to the same user; obtain other feature nodes connected to the target node in the overall relationship feature graph; use the target node and other feature nodes connected to the target node as matching feature nodes of user node k; Obtain the information vector of user node k and the information vectors of all user nodes connected to user node k, where the information is obtained by concatenating the attribute vector of the user node and the attribute vector of the edge; construct a node adjacent matrix based on the information vectors of user node k and all user nodes connected to user node k, where each information vector constitutes a row of the node adjacent matrix; For each matching feature node x, obtain a feature information vector of the matching feature node x, the feature information vector being composed of the node attribute feature information and the relationship feature information; and construct a feature adjacent matrix by taking all feature information vectors as rows of the feature adjacent matrix; The exchange entropy between the feature adjacency matrix and the node adjacency matrix is ​​used as the third attention index.

8. The user relationship mining system based on human networking according to claim 6 is characterized in that: If the loss function is greater than the first set value, and the first attention vector and the second attention vector meet the first condition, an attention adjustment matrix is ​​constructed based on the first attention vector, the second attention vector and the third attention vector, including: Obtaining matching point pairs, the matching point pairs including user nodes in the user interaction graph and feature nodes in the overall relationship feature graph that match each other; For each matching point pair, obtain the first cross entropy between the first attention vector of the user node and the second attention vector of the feature node; obtain multiple first cross entropies corresponding to the user interaction graph; obtain the average of the multiple first cross entropies; If the loss function is greater than the first set value, and the average of the multiple first cross entropies is greater than the second set value, for each user node, the first attention vector, the second attention vector, and the third attention vector corresponding to the user node are mapped to the same space to obtain a first mapping vector, a second mapping vector, and a third mapping vector, respectively; Performing a weighted summation on the first mapping vector, the second mapping vector, and the third mapping vector to obtain an attention adjustment vector of the user node; the attention adjustment vector includes a plurality of attention adjustment factors, each attention adjustment factor corresponding to an edge connected to the user node; Multiple attention adjustment vectors are arranged in sequence to form an attention adjustment matrix.

Images