User alignment method for heterogeneous e-commerce platforms based on meta-path-based full-view information representation

Abstract

The present invention belongs to the field of network alignment, and specifically relates to a method for aligning users of heterogeneous e-commerce platforms based on meta-path based full-view information representation, the method comprising: obtaining user attribute data and interaction data, and performing aggregation preprocessing on the user attribute data and interaction data; extracting the user data after aggregation preprocessing; inputting the extracted user data into a meta-path based full-view information representation heterogeneous e-commerce platform user alignment model to perform alignment processing on the users; the present invention starts from three aspects: user attribute information, user product interaction information, and user interest attention, wherein the key part of the invention is to use meta-path to aggregate sparse data of heterogeneous platforms, obtain more effective experimental data, and use attention mechanism to weight the learned user attribute relationship vector and user interaction relationship vector, so as to perform efficient alignment tasks.

Description

Technical Field

[0001] The present invention belongs to the field of network alignment, and in particular relates to a method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation. Background Art

[0002] With the popularization of mobile smart terminals and the rapid development of Internet technology, most people will create their own accounts on different platforms. However, due to the mutual exclusivity between platforms, it is difficult to determine whether two accounts on different platforms are used by the same user. In order to identify and match users across platforms, that is, to align heterogeneous network users, early researchers mainly proposed entity alignment methods based on user attribute information. This method measures the similarity between accounts and performs identification and matching based on the public profile information filled in by users, including nicknames, birthdays, addresses, gender, profile photos, etc. However, with the improvement of personal security awareness and the continuous improvement of platform privacy protection measures, the real identity information of most users will not be made public on the platform, so this method is not accurate. Then, an entity alignment method based on user structural relationships was proposed. This method uses the relationship structure between users and other nodes to align by comparing the similarity of their neighboring nodes.

[0003] In the field of cross-platform entity alignment, researchers have conducted extensive research on user attributes, user behaviors, and user relationship structures, and have achieved a lot of research results, but there are still some problems:

[0004] 1. The problem of sparse structural relations of heterogeneous platform network data nodes. There are many types of network data nodes on different e-commerce platforms, and the amount of data generated is very considerable, but the effective user data is relatively sparse, which makes it difficult to obtain key useful information of user nodes for user identity authentication.

[0005] 2. Diversity of data node relationships on heterogeneous platforms. E-commerce platforms have the characteristics of complex user attribute information and diverse user-product relationships. In addition, the relationship between platform data features is complex, and some user behaviors and relationships will not appear at the same time. How to integrate different node information and better represent the overall picture of the user information feature space has become a problem.

[0006] 3. How to integrate different relationships: The analysis of single personal attribute information or structural relationships is one-sided, and in actual systems, the interactive information is very rich, and it is difficult to integrate such information to identify the user's identity. Summary of the invention

[0007] In order to solve the problems existing in the above prior art, the present invention proposes a method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation, the method comprising:

[0008] S1. Extracting basic attribute data and historical interaction data of users of the first platform;

[0009] S2, using a meta-path-based entity-relationship aggregation model to aggregate the user's basic attribute data and historical interaction data to obtain aggregated user data;

[0010] S3, using the USRE representation model to represent the aggregated user data, and obtain the embedding vector of the user node user attribute information and the embedding vector of the user interaction information;

[0011] S4. The alignment model URSTA performs user matching on the embedding vectors of user attribute information and user interaction information, uses the similarity between cross-network user representations to determine the correspondence, and selects the user with the highest similarity in another platform for alignment.

[0012] Furthermore, the similarity between the cross-network user representations is used to determine the correspondence. The users of the first platform select the users with the highest similarity in the second platform for alignment. The similarity is expressed as:

[0013]

[0014] Among them, θ XY AB is the unified representation of the embedding vector of user attribute information and the embedding vector of user interaction information, U X A For users A,U in network X Y B is user B in network Y, n represents the total number of meta-path instances, X i is the vector representation of different meta-path instances.

[0015] Furthermore, θ XY AB is the unified representation of the embedding vector of user attribute information and the embedding vector of user interaction information, and its calculation formula is:

[0016]

[0017] in, represents the normalized importance weight of all meta-paths, UserA is user node A in network X, UserB is user node B in network Y, f(UserA, UserB) is the distance between user node A and user node B in the vector space obtained based on user attribute information, and g(UserA, UserB) is the distance between user node A and user node B in the vector space obtained based on user interaction information.

[0018] Furthermore, the alignment model URSTA introduces a multi-head attention mechanism to learn the normalized importance weights of all meta-paths The process includes:

[0019] Calculate the attention coefficients of all neighbor nodes of different meta-path instances;

[0020] Normalize the attention coefficients of all neighbor nodes through the softmax function;

[0021] The attention layer aggregates the embedding vectors of neighbor nodes of all relational meta-path types of the user node to obtain the normalized importance weights of all meta-paths.

[0022] Furthermore, f(UserA, UserB) is a node relationship group obtained based on user attribute information. A ,R UA ,V i >、<U B ,R UB ,V i >The calculated distance is calculated as follows: f(UserA, UserB) = ω AB |U A -U B | 2 ; g(UserA, UserB) is the node relationship group obtained based on user interaction information <U A ,R AI ,V i >、<U B ,R BI ,V i >The calculated distance is calculated as follows: g(UserA,UserB)=W AB ||U A +R i -U B || 2 ;

[0023] Among them, U A , U B and V i are the embedding vectors of user node A, user node B, and link node i, respectively. Link node i is the connection node shared by user node A and user node B. i For U A , U B and V i The connection relationship between them, R UA is the attribute information of user node A, R UB is the attribute information of user node B, R AI is the interaction information of user node A, RBI is the interaction information of user node A, ω AB represents the weight of user node A and user node B under the attribute relationship, W AB It represents the weight of user node A and user node B in the interaction relationship.

[0024] Furthermore, the process of aggregating the user's basic attribute data and historical interaction data includes:

[0025] S21, taking the meta-path as the context of the interaction between users and products, learning the representation of users, products and the interaction context between them, and aggregating the relationship information of all users;

[0026] S22, inputting the basic attribute data and historical interaction data of the current user into the entity-relationship aggregation model based on the meta-path;

[0027] S23. Using the user's interest and attention, the data output from the aggregation model is processed, and the processed data is used as aggregated user data.

[0028] Furthermore, the method of aggregating the basic attribute data of the user between meta-paths is as follows:

[0029]

[0030] The method of aggregating historical interaction data between meta-paths is:

[0031]

[0032] Among them, u i represents the user, r represents the path, n represents the total number of meta-path instances, and U(u i ) Attention represents the user's interest, P represents the meta-path instance of user data, <u i ,r j > indicates user u i In the current meta path instance r j The following relational context.

[0033] Furthermore, in combination with the user structure relationship, which includes user attribute information and user interaction information, the process of using the USRE representation model to represent the aggregated user data includes:

[0034] S31, using the USRE representation model to represent the overall information of the first platform network user structure relationship;

[0035] S32, obtaining a user attribute information node set and a user attribute information node set under different meta-path instances according to the overall information representation of the user structure relationship;

[0036] S33, obtaining a user interaction information node set and a user interaction information node set under different meta-path instances according to the overall information representation of the user structure relationship;

[0037] S34, synthesizing the four sets obtained in step S32 and step S33, obtaining the embedding vector of the user attribute information of the current user node and the embedding vector of the user interaction information.

[0038] Beneficial effects of the present invention:

[0039] In the feature relationship learning process, the present invention adopts an entity-relationship aggregation model based on meta-paths, obtains node information relationships through different meta-paths, and aggregates rich neighbor node information, solving the problem of sparse user structure relationship data.

[0040] Aiming at the characteristics of different structural relations, the present invention proposes a structural relation representation model of user attribute relation UARE and user interaction relation UIRE, which can not only distinguish the representation of different relation structure feature spaces in heterogeneous information networks, but also make full use of user overall feature information.

[0041] The present invention introduces the attention mechanism and proposes a user relationship sequence alignment model URSTA to perform integrated learning on user attribute relationships and user interaction relationships, and adopts a weight parameter allocation mechanism for user relationship features to better capture important influencing information.

[0042] Starting from the user attribute data and user interaction data of heterogeneous e-commerce platforms, the present invention proposes a meta-path-based full-view information representation heterogeneous e-commerce platform user alignment prediction model. This method considers three aspects: user attribute information, user product interaction information, and user interest attention. The key part of the invention is to use meta-paths to aggregate sparse data of heterogeneous platforms, obtain more effective experimental data, and use attention mechanisms to weight the learned user attribute relationship vectors and user interaction relationship vectors, so as to perform efficient alignment tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is an overall block diagram of the present invention;

[0044] Figure 2 This is an example of user relationship data aggregation in the present invention;

[0045] Figure 3 Schematic diagram of user relationship attention weight allocation in the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] A user alignment method for heterogeneous e-commerce platforms based on meta-path full picture information representation, such as Figure 1 As shown, the method includes:

[0048] Obtaining the user's attribute data and interaction data, and performing aggregation preprocessing on the user's attribute data and interaction data;

[0049] Extract relevant attribute relationships and interaction relationships of user data after aggregation preprocessing;

[0050] The relevant attribute relationships and interaction relationships of the extracted user data are input into the meta-path-based full-view information representation heterogeneous e-commerce platform user alignment model to align the users.

[0051] In another embodiment, the method comprises:

[0052] S1, extracting basic attribute data and historical interaction data of users of the bookstore e-commerce platform;

[0053] S2, using a meta-path-based entity-relationship aggregation model to aggregate the user's basic attribute data and historical interaction data to obtain aggregated user data;

[0054] S3, using the USRE representation model to represent the aggregated user data, and obtain the embedding vector of the user node user attribute information and the embedding vector of the user interaction information;

[0055] S4. Use the alignment model URSTA to perform user matching on the embedding vectors including user attribute information and the embedding vectors including user interaction information, and select the user with the highest similarity in the other platform for alignment.

[0056] The process of data aggregation for sparse user row data in step S2 is as follows:

[0057] Taking the meta-path as the context of the interaction between users and products, we learn the representation of users, products and the context of their interaction, and aggregate the relationship information of all users.

[0058] Input the basic attribute data and historical interaction data of the current user into the entity-relationship aggregation model based on the meta-path;

[0059] The data output from the aggregation model is processed using user interest and attention, and the processed data is used as aggregated user data.

[0060] The method of aggregating the basic attribute data of users between meta-paths is to use the aggregation model to pre-process the basic attribute data of users, which is expressed as:

[0061]

[0062] The method of aggregating historical interaction data between meta-paths is to use an aggregation model to pre-process the user's historical interaction data, which can be expressed as:

[0063]

[0064] Among them, u i represents the user, r represents the path, n represents the total number of meta-path instances, and U(u i ) Attention represents the user's interest, P represents the meta-path instance of user data, <u i ,r j >Indicates the different relational contexts of the user under the current meta-path instance.

[0065] In order to distinguish the differences between different attribute relationships, the process of using the USRE representation model to represent the aggregated data in step S3 includes:

[0066] S31. Denote the bookstore e-commerce network as G x =(V x ,F x ,E x ), using the USRE representation model to represent the overall information of the user structure relationship of the bookstore e-commerce network;

[0067] Among them, x represents the platform, V is the node set, F is the link relationship, and E is the edge;

[0068] S32, obtaining a user attribute information node set and a user attribute information node set under different meta-path instances according to the overall information representation of the user structure relationship;

[0069] S33, obtaining a user interaction information node set and a user interaction information node set under different meta-path instances according to the overall information representation of the user structure relationship;

[0070] S34, synthesizing the four sets obtained in step S32 and step S33, obtaining an embedding vector including user attribute information and an embedding vector including user interaction information of the current user node.

[0071] In the process of the USRE representation model representing the aggregated data, its user structural relationship includes all user attribute information and all user interaction information.

[0072] For user attribute relationships, the nodes connected by user attributes are information that the user has, so the nodes are very close in the vector space, which conforms to the optimization of the Euclidean distance. For attribute features, it is necessary to first vectorize multiple user attribute information, which becomes the key to calculating the similarity of user attribute information. After extracting the basic attribute data of the user, it is processed, that is, the vector of the user attribute information is projected into the same latent factor space, and a specific type of linear transformation is applied to each type of node, which is expressed as:

[0073]

[0074] Among them, X v is the vector of user attribute information, Z v is the new latent feature vector, θ Y is the parameter weight matrix of the Y-type node.

[0075] In the alignment model URSTA, a multi-head attention mechanism is introduced to distribute the weights of user attribute information and user interaction information, aiming to distinguish the different importance of different attribute features of a user in the alignment task. Different meta-path instances will contribute to the representation of user nodes to different degrees, including:

[0076] Calculate the attention coefficients of all neighbor nodes of different meta-path instances:

[0077] In order to facilitate the comparison of different attention coefficients, the attention coefficients of all neighbor nodes are normalized by the softmax function:

[0078] The attention layer aggregates the embedding vectors of neighbor nodes of all meta-path types P of user node u to obtain the normalized importance weights:

[0079] Among them, P represents different meta-path instances, ReLU is the first activation function, is a hyperparameter, h p (u,v) is the meta-path, σ is the second activation function, Indicates the importance of the current path instance of user node u, It represents the sum of weights of all meta-path instances under the current meta-path of user node u. It represents the normalized importance weight of all meta-paths of the user node under the attribute relationship, that is, the weighted combination of all meta-paths of the user node u.

[0080] The representation power of the attention layer is improved by leveraging a multi-head attention mechanism, where multiple attention units operate on different low-dimensional projections of the input, namely each meta-path in this paper.

[0081] In order to distinguish the different weights of user attribute relationship EARE and user interaction relationship EIRE in alignment, a multi-parameter vector is set for weight learning. The standardized importance weight of each meta-path instance is learned, and then all instances are weighted summed for modeling. The formula is as follows:

[0082] Objective function:

[0083] Similarity calculation formula:

[0084] in, represents the normalized importance weight of all meta-paths of user nodes under attribute relations, represents the normalized importance weight of all meta-paths of user nodes under the interaction relationship, θ XY AB is the unified representation of the embedding vector of user attribute information and the embedding vector of user interaction information, U X A For users A,U in network X Y B is user B in network Y, n represents the total number of meta-path instances, X i are vector representations of different meta-path instances, UserA is user node A in network X, UserB is user node B in network Y, f(UserA, UserB) is the distance between user node A and user node B in the vector space obtained based on user attribute information, and g(UserA, UserB) is the distance between user node A and user node B in the vector space obtained based on user interaction information.

[0085] Preferably, f(UserA, UserB) is a node relationship group obtained according to user attribute information. A ,R UA ,V i >、<U B ,R UB ,V i >The calculated distance is calculated as follows: f(UserA, UserB) = ω AB |U A -U B | 2 ;

[0086] g(UserA, UserB) is the node relationship group obtained based on user interaction information.A ,R AI ,V i >、<U B ,R BI ,V i >The calculated distance is calculated as follows: g(UserA,UserB)=W AB ||U A +R i -U B || 2 ;

[0087] Among them, U A , U B and V i are the embedding vectors of user node A, user node B, and link node i, respectively. Link node i is the connection node shared by user node A and user node B. i For U A , U B and V i The connection relationship between them, R UA is the attribute information of user node A, R UB is the attribute information of user node B, R AI is the interaction information of user node A, R BI is the interaction information of user node A, ω AB represents the weight of user node A and user node B under the attribute relationship, W AB It represents the weight of user node A and user node B in the interaction relationship.

[0088] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation, characterized in that: The method comprises the following steps: S1. Extracting basic attribute data and historical interaction data of users of the first platform; S2, using a meta-path-based entity-relationship aggregation model to aggregate the user's basic attribute data and historical interaction data to obtain aggregated user data; The process of aggregating the user's basic attribute data and historical interaction data includes: S21, taking the meta-path as the context of the interaction between users and products, learning the representation of users, products and the interaction context between them, and aggregating the relationship information of all users; S22, inputting the basic attribute data and historical interaction data of the current user into the entity-relationship aggregation model based on the meta-path; S23, using the user's interest attention, processing the data output from the aggregation model, and using the processed data as aggregated user data; The method of aggregating the basic attribute data of users between meta-paths is as follows: The method of aggregating historical interaction data between meta-paths is: Among them, u i represents the user, r represents the path, n represents the total number of meta-path instances, and U(u i ) Attention represents the user's interest, P represents the meta-path instance of user data, <u i ,r j > indicates user u i In the current meta path instance r j The relational context below; S3, using the USRE representation model to represent the aggregated user data, and obtain the embedding vector of the user node user attribute information and the embedding vector of the user interaction information; S4. The alignment model URSTA performs user matching on the embedding vectors of user attribute information and user interaction information, uses the similarity between cross-network user representations to determine the correspondence, and selects the user with the highest similarity in another platform for alignment.

2. According to claim 1, the method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation is characterized in that: The correspondence is determined by using the similarity between user representations across networks. Users on the first platform select users with the highest similarity on the second platform for alignment. The similarity calculation formula is: Among them, θ XY AB is the unified representation of the embedding vector of user attribute information and the embedding vector of user interaction information, U X A For users A,U in network X Y B is user B in network Y, n represents the total number of meta-path instances, X i is the vector representation of different meta-path instances.

3. The method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation according to claim 2 is characterized in that: θ XY AB is the unified representation of the embedding vector of user attribute information and the embedding vector of user interaction information, and its calculation formula is: in, represents the normalized importance weight of all meta-paths, UserA is user node A in network X, UserB is user node B in network Y, f(UserA, UserB) is the distance between user node A and user node B in the vector space obtained based on user attribute information, and g(UserA, UserB) is the distance between user node A and user node B in the vector space obtained based on user interaction information.

4. The method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation according to claim 1 or 3, characterized in that: The alignment model URSTA introduces a multi-head attention mechanism to learn the normalized importance weights of all meta-paths The process includes: Calculate the attention coefficients of all neighbor nodes of different meta-path instances; Normalize the attention coefficients of all neighbor nodes through the softmax function; The attention layer aggregates the embedding vectors of neighbor nodes of all meta-path types of the user node to obtain the normalized importance weights of all meta-paths 5. According to claim 3, the method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation is characterized in that: f(UserA, UserB) is the node relationship group obtained based on user attribute information <U A ,R UA ,V i >、<U B ,R UB ,V i >The calculated distance is calculated as follows: f(UserA,UserB) = ω AB |U A -U B | 2 ; g(UserA, UserB) is the node relationship group obtained based on user interaction information. A ,R AI ,V i >、<U B ,R BI ,V i >The calculated distance is calculated as follows: g(UserA,UserB)=W AB ||U A +R i -U B || 2 ; Among them, U A , U B and V i are the embedding vectors of user node A, user node B, and link node i, respectively. Link node i is the connection node shared by user node A and user node B. i For U A , U B and V i The connection relationship between them, R UA is the attribute information of user node A, R UB is the attribute information of user node B, R AI is the interaction information of user node A, R BI is the interaction information of user node A, ω AB represents the weight of user node A and user node B under the attribute relationship, W AB It represents the weight of user node A and user node B in the interaction relationship.

6. The method for aligning users of heterogeneous e-commerce platforms based on meta-path-based full-view information representation according to claim 1 is characterized in that: Combined with the user structure relationship, which includes user attribute information and user interaction information, the process of using the USRE representation model to represent the aggregated user data includes: S31, using the USRE representation model to represent the overall information of the first platform network user structure relationship; S32, obtaining a user attribute information node set and a user attribute information node set under different meta-path instances according to the overall information representation of the user structure relationship; S33, obtaining a user interaction information node set and a user interaction information node set under different meta-path instances according to the overall information representation of the user structure relationship; S34, synthesizing the four sets obtained in step S32 and step S33, obtaining the embedding vector of the user attribute information of the current user node and the embedding vector of the user interaction information.

Images