A cross-e-commerce bookstore user alignment method based on tensor decomposition and global representation
Through tensor decomposition and full-view representation methods, combined with the UBN2vec model and interest group block index model, the problems of sparse user behavior data and complex full-view data on e-commerce platforms are solved, and accurate and efficient alignment of cross-platform users is achieved, reducing the computational complexity.
Patent Information
- Application Number
- CN202211408791.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-11
- Publication Date
- 2025-08-29
- Estimated Expiration
- 2042-11-11
AI Technical Summary
The effective user behavior data in e-commerce platforms is sparse, the full data feature space of heterogeneous e-commerce platforms is complex, and the cross-platform user matching process is computationally complex, and it is difficult for existing methods to achieve accurate and efficient user alignment.
Using a method based on tensor decomposition and full-view representation, data compensation is performed through the ‘user-behavior-book’ tensor decomposition model, user data processing is performed using the UBN2vec model, and user interest groups are divided through the interest group block index model, and finally the similarity of user node embedding representation vectors in the interest group is calculated to match.
It realizes accurate and efficient cross-platform user alignment on e-commerce platforms, reduces the computational complexity, reduces the number of matches for users with greater interests, and improves the efficiency of the matching process.
Smart Images
Figure CN115907865B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of network alignment, and in particular relates to a cross-e-commerce bookstore user alignment method based on tensor decomposition and global representation. Background Art
[0002] In recent years, with the rapid development of internet technology and the advent of the big data era, various social networks and e-commerce platforms have emerged. To meet diverse needs, people register accounts on multiple social networks and e-commerce platforms. Heterogeneous network user alignment technology aims to identify and match users distributed across different platforms. In the commercial sector, user alignment technology facilitates cross-platform recommendations and information dissemination. In national security, user alignment technology clears obstacles for important tasks such as anonymous user discovery and public opinion analysis. Consequently, research on heterogeneous network alignment has attracted widespread attention from scholars both domestically and internationally.
[0003] Researchers both domestically and internationally have conducted extensive research on user alignment technology. Existing user alignment methods can be broadly categorized into two main categories: those based on user attributes and structure, and those based on machine learning and deep learning. These methods utilize the attribute information publicly available on social platforms and the topological structure of social networks to measure similarity between users. Machine learning and deep learning methods manually extract user features through feature engineering, train classification models based on labeled data, and then use the trained models to perform cross-platform user alignment predictions.
[0004] At present, although a lot of research has been done on heterogeneous network alignment technology and fruitful research results have been achieved, some problems still exist:
[0005] 1. Valid user behavior data on e-commerce platforms is sparse: Although the total amount of data generated by users on e-commerce platforms is considerable, valid user behavior information is scarce. Therefore, the difficulty in obtaining valid user information poses a challenge to cross-platform user identity matching.
[0006] 2. The complexity of the feature space of heterogeneous e-commerce platform data: Due to the diverse node types and complex user behaviors in heterogeneous e-commerce platforms, how to effectively represent the feature space of the entire e-commerce platform affects the accuracy of alignment prediction;
[0007] 3. High computational complexity of the cross-platform user matching process: E-commerce platforms have a large number of user entities. The computational complexity and time consumption of identifying and matching massive numbers of users across large e-commerce platforms is enormous. Summary of the Invention
[0008] To solve the above technical problems, the present invention proposes a method for aligning users of cross-e-commerce bookstores based on tensor decomposition and global representation, comprising:
[0009] S1: Extract basic attribute information and user behavior data of users on the e-commerce bookstore platform;
[0010] S2: Use the "user-behavior-book" tensor decomposition model to perform homomorphic compensation on user behavior data information to obtain compensated user behavior data;
[0011] S3: Use the UBN2vec global representation model to process the compensated e-commerce bookstore user data to obtain a user node embedding representation vector containing rich attribute information, behavioral preference information, and user neighborhood information;
[0012] S4: Use the user interest group block index model to divide the e-commerce bookstore users into interest groups, and divide users with different interest preferences into different interest groups;
[0013] S5: Calculate the similarity between the embedding representation vectors of user nodes in the same interest group, and select the user with the highest similarity in the other platform as the alignment user to be matched.
[0014] The beneficial effects of the present invention are as follows: the present invention utilizes the characteristics of the tensor decomposition model in data conversion, iteratively processes missing values during the tensor decomposition process, performs homomorphic compensation for the sparse data of the e-commerce platform, and utilizes the representation learning model to low-rank dense vectorize the feature space of the complex heterogeneous e-commerce bookstore platform. Finally, the attribute information and behavioral data information of the users in the bookstore platform are comprehensively considered to achieve accurate cross-platform user alignment, reduce the number of matches for users with large differences in interests, reduce the complexity of the matching process, and finally accurately and efficiently achieve alignment of users across bookstore platforms. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 is a flow chart of the method of the present invention;
[0016] Figure 2 Schematic diagram of the "user-behavior-book" tensor decomposition model of the present invention;
[0017] Figure 3 Schematic diagram of multi-viewing of user behavior in the UBN2vec model of the present invention;
[0018] Figure 4 This is a schematic diagram of the process of dividing the abstract book types into interest categories according to the present invention;
[0019] Figure 5 Schematic diagram of user interest groups according to the present invention. DETAILED DESCRIPTION
[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.
[0021] A cross-e-commerce bookstore user alignment method based on tensor decomposition and full-image representation, such as Figure 1 As shown, including:
[0022] S1: Extract basic attribute information and user behavior data of users on the e-commerce bookstore platform;
[0023] S2: Use the "user-behavior-book" tensor decomposition model to perform homomorphic compensation on user behavior data information to obtain compensated user behavior data;
[0024] S3: Use the UBN2vec global representation model to process the compensated e-commerce bookstore user data to obtain a user node embedding representation vector containing rich attribute information, behavioral preference information, and user neighborhood information;
[0025] S4: Use the user interest group block index model to divide the e-commerce bookstore users into interest groups, and divide users with different interest preferences into different interest groups;
[0026] S5: Calculate the similarity between the embedding representation vectors of user nodes in the same interest group, and select the user with the highest similarity in the other platform as the alignment user to be matched.
[0027] Obtaining user data involves directly accessing data sources provided by the bookstore, downloading existing public data sources, or obtaining raw data through the e-commerce network's public application programming interface (API). Since raw data is unstructured, it cannot be directly used for user data analysis; therefore, data preprocessing is required to structure the largely unstructured data. This preprocessing process includes deduplication and removing invalid nodes, such as visitor nodes that do not contain any behavioral information.
[0028] The process of building the "user-behavior-book" tensor decomposition model includes:
[0029] S21: Obtain user historical behavior information and construct a tensor M∈R with the dimension of "user-behavior-book" to be completed I ×J×K ;
[0030] S22: Treat the completed tensor M through tensor decomposition, and restore the complete low-rank tensor Y∈R based on the partial user behavior information observed in the observable index set Ω I×J×K ; The process is expressed as:
[0031]
[0032] Among them, ‖Y‖ * Represents the trace norm of the tensor Y.
[0033] S23: Decompose the tensor into the model's tensor trace norm ‖Y‖ * Convert all matrices y that are expanded along n modes of the tensor Y (i) The convex combination of the trace norm of , results in a convex optimization model; the model is expressed as:
[0034]
[0035] Among them, ω i is the combined coefficient of the matrix trace norm, ω i Satisfy i ≥0 and y (i) Represents the matrix of tensor Y expanded along n modes; ||y (i) || * Represents the matrix y (i) The trace norm of .
[0036] S24: Use high-precision low-rank tensor completion algorithm to solve convex optimization model, and introduce n auxiliary variables during the algorithm operation An equivalent convex optimization model is obtained; the model is expressed as:
[0037]
[0038] Among them, A i(i) is the auxiliary tensor A i The matrix expanded along n modes, ||A i(i) || * is the matrix A i(i) The trace norm of .
[0039] S25: Integrate the augmented Lagrangian function into the equivalent convex optimization model to obtain the “user-behavior-book” tensor decomposition model; the model is expressed as:
[0040]
[0041] Among them, Y is the new tensor that is iteratively compensated by the original tensor M, is the set of Lagrange multiplier tensors, λ is the penalty factor, A is the set of auxiliary tensors introduced in S24, A i(i) is the auxiliary tensor Ai The matrix expanded along n modes, ||A i(i) || * is the matrix A i(i) The trace norm of . <Y-A i ,W i > represents the tensor YA i and tensor W i The inner product of Represents the squared value of the F-norm of the difference between each introduced intermediate variable and the complete tensor of user behavior.
[0042] like Figure 2 As shown in the figure, the “user-behavior-book” tensor decomposition model is used to perform homomorphic compensation on sparse user behavior data, including:
[0043] The ADMM algorithm is used to solve the extreme value of the cost function of the equivalent convex optimization model and iteratively update A i , Y and W i , obtain the user behavior data after compensation;
[0044] A i The update formula is:
[0045]
[0046] The update formula of Y is:
[0047]
[0048] W i The update formula is:
[0049]
[0050] Among them, argminL(.) represents the variable value when the objective function reaches the minimum value, A k+1 Indicates the value of the auxiliary tensor A at the k+1th iteration, Y k+1 Represents the value of tensor Y at the k+1th iteration, W i k+1 Indicates W at the k+1th iteration i The value of ,λ represents the penalty factor, fold n (.) means reconstructing the n pattern matrices of the tensor into a tensor, Represents the singular value contraction operation, the contraction operator is W i / λ,Y (i) Represents the i-mode matrix of the tensor Y, W i(i) represents the Lagrange multiplier tensor W i The i-mode matrix.
[0051] In order to effectively represent the overall feature space of the heterogeneous e-commerce bookstore platform, the UBN2vec model is used to independently embed the three elements that constitute the heterogeneous bookstore e-commerce network: user attributes, user behavior, and user neighborhood topology structure. The independent embedding vectors are then fully fused to obtain the final low-rank dense overall embedding vector, thereby comprehensively considering the information of all the constituent elements in the heterogeneous e-commerce bookstore network.
[0052] The UBN2vec full-view representation model includes: user neighborhood embedding module, user behavior embedding module, and user attribute embedding module;
[0053] A heterogeneous e-commerce bookstore network G = (V, E) is constructed based on the actual behavior records of e-commerce bookstore users. The deepwalk model is used to learn the network topology of the heterogeneous e-commerce bookstore network G = (V, E). The node attribute information and user behavior types of the heterogeneous e-commerce bookstore network are ignored, and the heterogeneous e-commerce bookstore network is simplified into a homogeneous network with a single edge type, thereby obtaining a user neighborhood embedding module. According to the different edge types in the heterogeneous e-commerce bookstore network G = (V, E), the original e-commerce bookstore network is separated into several sub-networks containing only a certain specific edge type, thereby obtaining a user behavior embedding module. According to the attribute encoding and attribute feature conversion in the heterogeneous e-commerce bookstore network G = (V, E), a user attribute embedding module is obtained. The user neighborhood embedding module, the user behavior embedding module, and the user attribute embedding module are integrated to obtain the UBN2vec full-view representation model.
[0054] The UBN2vec global representation model is used to process the compensated e-commerce bookstore user data, including:
[0055] S31: Randomly select a user node or book node in the e-commerce bookstore network as the starting node v i , from node v i Start by generating m truncated random walk sequences W with a maximum length of k (i) ;
[0056] S32: Adopt the skip-gram model strategy in word2vec, according to the central node v in each random walk sequence m The embedding representation of the predicted surrounding nodes v j Embedding representation, establish the likelihood probability function and mapping function of node prediction;
[0057]
[0058] Among them, E represents a mapping function or can be understood as the vector representation matrix X of all nodes in the bookstore platform |V|×d(The matrix dimension is |V|×d, |V| represents the number of nodes in the bookstore platform, and d represents the vector dimension of each node), which can map each node in the bookstore platform to a d-dimensional vector. E(v m ) represents node v m The embedding vector, P(v j |E(v m )) indicates that node v m Embedding vector of the surrounding nodes v j The probability function of the embedding vector, P({v m-k / 2 ,…,v m+k / 2}\v m |E(v m )) indicates that node v m The embedding vector predicts the random walk sequence except v m The likelihood probability function of the embedding vector of other nodes except m-(k / 2) Represents the node with subscript m-(k / 2) in the random walk sequence of length k.
[0059] Furthermore, the probability function P(v j |E(v m ))'s expression is as follows;
[0060]
[0061] Among them, E(v j ) T Represents node v j The embedding vector E(v j )’s transposed vector, E(v j ) T ·E(v m ) represents node v m and node v j The dot product of the embedding vector; E(v n ) T Represents node v n The embedding vector E(v n ), E(v n ) T ·E(v m ) represents node v m and node v n Dot product of embedding vectors.
[0062] S33: Iteratively update each parameter in the mapping function E so that the negative log-likelihood function converges, and the parameters in the mapping function finally obtained are the final neighborhood embedding vectors of each node;
[0063]
[0064] Among them, E represents the mapping function, that is, the vector representation matrix X of all nodes in the bookstore platform |V|×d , P({v m-(k / 2) ,…,v m+(k / 2)}\v m |E(v m )) indicates that node v m The embedding vector predicts the random walk sequence except v m The likelihood probability function of the embedding vector of other nodes except m ) represents node v m The embedding vector, v m-(k / 2) Represents the node with subscript m-(k / 2) in the random walk sequence of length k.
[0065] S34: Based on the four types of user behaviors in the bookstore network: purchase, add to cart, click, and friendship between users, it is divided into four sub-views: "buy", "collect", "click", and "friend";
[0066] User behavior is multi-viewed, separating different types of user behavior edges into different behavior views; Figure 3 As shown in the figure, there are four types of user behaviors in the e-commerce bookstore network: purchase, add to cart, click, and friendship between users. This is divided into four sub-views: "buy," "collect," "click," and "friend." Each sub-view only contains specific user behavior edge types and only considers behavior pattern information of that specific type.
[0067] S35: Each user node u i The embedding representation vectors for specific user behaviors are generated independently in different behavior views; this process is expressed as:
[0068]
[0069] in, Represents user node u i When the user behavior type is behavior (i) The aggregate vector of the k-th order neighbors on the behavioral view of Represents user node u i The aggregation vector of the k-1th order neighbors, Indicates that the user behavior type is behavior (i) The set of neighbor nodes in the view of Represents the aggregation vector The neighbor node element v contained in j Belongs to user node u i The set of neighbor nodes mean(.) represents the mean aggregation function, and σ is the activation function.
[0070] S36: User node u i The embedding vectors under each subview are concatenated into the overall behavior embedding vector;
[0071] S37: Generate user attribute embedding vector u i(attr) ; The process is expressed as:
[0072]
[0073] in, is a human-defined hyperparameter, D Z is the feature transformation matrix for different types of nodes, x i is node u i Initialization attribute feature vector
[0074] S38: Fuse the user neighborhood embedding vector, user behavior embedding vector, and user attribute embedding vector into the final overall embedding vector; this process is expressed as:
[0075]
[0076] in, is the user node u i In user behavior type behav (i) The overall embedding vector u contains user neighborhood information, user behavior information, and user attribute information. i(neighbor) is the user node u i The neighborhood embedding vector of Is the behavior type of behavior (i) The importance coefficient of user behavior embedding relative to the overall embedding, is the feature transformation matrix, Represents the transposed matrix of the feature transformation matrix, u i(behavior) Represents user node u i The overall behavior embedding vector, It is the user behavior type calculated based on the self-attention mechanism (i) Next, embed the vector u for the overall behavior i(behavior) The linear combination coefficients of .
[0077] Further, The calculation formula is as follows:
[0078]
[0079] Among them, W beha(i) and w behav(i)It is a trainable matrix of different dimensions under user behavior type behav(i).
[0080] Traditional user alignment research requires similarity matching between user nodes in the target network and all user nodes in the target network, ultimately selecting the optimal set of matching pairs. This round-robin matching process is computationally extremely complex. To reduce this computational complexity, a block indexing model based on user interest groups was constructed. This assigns users with similar interests to the same interest group, reducing the number of matches for users with significantly different interests and ultimately lowering the computational complexity of the user matching process.
[0081] User interest group block indexing model: Books are divided into multiple user interest groups based on the user's historical behavior information, and each user is assigned to a user interest group based on the user's actual purchase record.
[0082] Since the same user across platforms will show similar interests and preferences in different bookstore platforms, although the book categories purchased on the two platforms may not be completely the same, there is a certain degree of correlation between these book categories. Therefore, the present invention is based on each user's real purchase record and each user's interest in book types. Aggregate the book subcategories with high correlation into abstract book type interest categories. If the book purchased by the user belongs to a certain abstract book type interest category, the user will be assigned to the user interest group corresponding to the abstract book type interest category; the schematic diagram of the construction of abstract book type interest category is as follows Figure 4 shown.
[0083] Use the user interest group block index model to divide e-commerce bookstore users into interest groups. Divide e-commerce bookstore users into interest groups, including:
[0084] S41: Construct an initial book category association matrix to quantify the degree of association between each book subcategory. The initial book association matrix is an all-zero matrix.
[0085] S42: Calculate the correlation coefficient between the two book categories A and B, and compare it with the user u i The correlation coefficients between all related book subcategories are filled into the book category correlation matrix to obtain the final correlation matrix;
[0086] S43: Extract each column in the final correlation matrix, aggregate the top k book subcategories with the highest correlation in each column into abstract book interest categories, and assign each user to each abstract book interest category to form a user interest group based on the user's actual purchase record, such as Figure 5 shown.
[0087] Calculate the correlation coefficient between two book subcategories A and B, expressed as:
[0088]
[0089] Among them, R (type(A),type(B)) represents the correlation coefficient between two book subcategories A and B, Represents user u i The degree of interest in a certain subcategory of books. type(A) indicates subcategory A, and type(B) indicates subcategory B.
[0090] Assuming that a user's interest in book categories A, B, and C are 0.8, 0.5, and 0.1 respectively, the correlation coefficients between these book categories are as follows:
[0091]
[0092]
[0093]
[0094] and will be with user u i The correlation coefficients between all relevant book subcategories are added to the book category correlation matrix. This process is repeated until all the book subcategory correlation coefficients calculated based on each user's behavior information have been accumulated into the book category correlation matrix to obtain the final correlation matrix.
[0095] Useru i Calculation of interest in a certain book category:
[0096]
[0097] in, Represents user u i For the category type(b i )'s interest in the book, Represents user u i The book collection you purchased, The book category that the user purchased is type(b i ), |·| represents the number of elements in a collection, IsBuy represents whether the user has purchased, IsCollect represents whether the book has been collected, and IsClick represents whether the user has clicked. The values of IsBuy, IsCollect, and IsClick are 1, 0.8, and 0.6 respectively. Represents user u i For type type(b i ) is the total number of rows of books, and the symbol |.| represents the number of elements in a set.
[0098] Calculate the similarity between the embedding vectors of user nodes in the same interest group, including:
[0099]
[0100] in, represents the embedding vector of user 1 in the bookstore platform x, represents the embedding vector of user 2 in the bookstore platform y, The specific form is S1 represents vector The modulus of S2 represents the vector The modulus length, s 1i Represents a vector The i-th element in s 2i Represents a vector The i-th element in n represents the vector or The number of elements in , The specific form is represents the cosine similarity between the embedding vectors of user 1 and user 2.
[0101] By calculating the cosine similarity between the embedding vectors of bookstore users, the user with the highest similarity is used as the alignment user to be matched.
[0102] While 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 these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A cross-e-commerce bookstore user alignment method based on tensor decomposition and holistic representation, characterized by: include: S1: Extract basic attribute information and user behavior data of users on the e-commerce bookstore platform; S2: Use the "user-behavior-book" tensor decomposition model to perform homomorphic compensation on user behavior data information to obtain compensated user behavior data; The "user-behavior-book" tensor decomposition model is used to perform homomorphic compensation for sparse user behavior data, including: Among them, Y k+1 represents the value of the low-rank tensor Y at the k+1th iteration, that is, the user behavior data after homomorphic compensation, argminL(.) represents the variable value when the objective function achieves the minimum value, A k+1 represents the value of the auxiliary tensor A at the k+1th iteration, W i k+1 Represents the Lagrange multiplier tensor W at the k+1th iteration i The value of , λ represents the penalty factor; S3: Use the UBN2vec global representation model to process the compensated e-commerce bookstore user data to obtain a user node embedding representation vector containing rich attribute information, behavioral preference information, and user neighborhood information; The UBN2vec full-view representation model includes: user neighborhood embedding module, user behavior embedding module, and user attribute embedding module; Based on the actual behavior records of e-commerce bookstore users, a heterogeneous e-commerce bookstore network is constructed. The network topology of the heterogeneous e-commerce bookstore network is learned using the deepwalk model. The node attribute information and user behavior types of the heterogeneous e-commerce bookstore network are ignored, and the heterogeneous e-commerce bookstore network is simplified into a homogeneous network with a single edge type, resulting in a user neighborhood embedding module. Based on the different edge types in the heterogeneous e-commerce bookstore network, the original e-commerce bookstore network is separated into sub-networks containing edge types, resulting in a user behavior embedding module. Based on the integration of attribute encoding and attribute feature conversion into the heterogeneous e-commerce bookstore network, a user attribute embedding module is obtained. The user neighborhood embedding module, user behavior embedding module, and user attribute embedding module are integrated to obtain the UBN2vec full-view representation model. S4: Use the user interest group block index model to divide the e-commerce bookstore users into interest groups, and divide users with different interest preferences into different interest groups; User interest group block indexing model: Books are divided into multiple user interest groups based on the user's historical behavior information, and each user is assigned to a user interest group based on the user's actual purchase history; S5: Calculate the similarity between the embedding representation vectors of user nodes in the same interest group, and select the user with the highest similarity in the other platform as the alignment user to be matched.
2. A cross-e-commerce bookstore user alignment method based on tensor decomposition and holistic representation according to claim 1, characterized in that: The process of building the "user-behavior-book" tensor decomposition model includes: S21: Obtain user historical behavior information and construct a tensor M∈R with the dimension of "user-behavior-book" to be completed I×J×K , where R I×J×K Represents a third-order tensor, I, J, and K represent the three dimensions of the tensor R respectively; S22: Treat the completed tensor M through tensor decomposition, and restore the complete low-rank tensor Y∈R based on the partial user behavior information observed in the observable index set Ω I×J×K ; The process is expressed as: Subject to Y Ω =M Ω Among them, ‖Y‖ * represents the trace norm of the low-rank tensor Y, Y Ω Represents the element corresponding to the observable index set Ω in the tensor Y, M Ω Represents the element corresponding to the observable index set Ω in the tensor M; S23: Decompose the tensor into the model's tensor trace norm ‖Y‖ * Convert all matrices y that are expanded along n modes of the tensor Y (i) The convex combination of the trace norm of , results in a convex optimization model; the model is expressed as: Subject to Y Ω =M Ω Among them, ω i represents the combined coefficient of the matrix trace norm, y (i) Represents the matrix of the low-rank tensor Y expanded along n modes, ||y (i) || * Represents the matrix y (i) The trace norm of Y Ω Represents the element corresponding to the observable index set Ω in the tensor Y, M Ω Represents the element corresponding to the observable index set Ω in the tensor M; S24: Use high-precision low-rank tensor completion algorithm to solve convex optimization model, and introduce n auxiliary variables during the algorithm operation An equivalent convex optimization model is obtained; the model is expressed as: Among them, A i(i) is the auxiliary tensor A i The matrix expanded along n modes, ||A i(i) || * is the matrix A i(i) The trace norm of Y Ω Represents the element corresponding to the observable index set Ω in the tensor Y, M Ω Represents the element corresponding to the observable index set Ω in the tensor M, Y represents the low-rank tensor, ω i represents the combined coefficients of the matrix trace norm; S25: Integrate the augmented Lagrangian function into the equivalent convex optimization model to obtain the "user-behavior-book" tensor decomposition model; the model is expressed as: Among them, Y represents a low-rank tensor, represents the set of Lagrange multiplier tensors, λ is the penalty factor, A is the auxiliary tensor, A i(i) is the auxiliary tensor A i The matrix expanded along n modes, ||A i(i) || * is the matrix A i(i) The trace norm of <Y-A i ,W i > represents the tensor YA i and the Lagrange multiplier tensor W i The inner product of Represents the squared value of the F-norm of the difference between each introduced intermediate variable and the complete tensor of user behavior.
3. The method for cross-e-commerce bookstore user alignment based on tensor decomposition and holistic representation according to claim 1 is characterized in that: The UBN2vec global representation model is used to process the compensated e-commerce bookstore user data, including: S31: Randomly select a user node or book node in the e-commerce bookstore network as the starting node v i , from node v i Start by generating m truncated random walk sequences W with a maximum length of k (i) ; S32: Adopt the skip-gram model strategy in word2vec, according to the central node v in each random walk sequence m The embedding representation of the predicted surrounding nodes v j Embedding representation, establish the likelihood probability function and mapping function of node prediction; Likelihood probability function of node prediction: Mapping function: E:v∈V→X |V|×d Among them, E represents the mapping function, that is, the vector representation matrix X of all nodes in the bookstore platform |V|×d , |V|×d represents the matrix dimension, |V| represents the number of nodes in the bookstore platform, d represents the vector dimension of each node, E(v m ) means that the node v m Mapped to an embedded vector, that is, from the overall vector matrix X |V|×d Take out node v m The embedding vector, P(v j |E(v m )) indicates that node v m Predict surrounding nodes v j The probability function of the embedding representation, P({v m-(k / 2) ,…,v m+(k / 2) }\v m |E(v m )) indicates that node v m The embedding vector predicts the random walk sequence except v m The likelihood probability function of the embedding vector of other nodes except m-(k / 2) represents the node with subscript m-(k / 2) in the random walk sequence of length k; S33: Iteratively update each parameter in the mapping function E so that the negative log-likelihood function converges, and the parameters in the mapping function finally obtained are the final neighborhood embedding vectors of each node; Among them, E represents the mapping function, that is, the vector representation matrix X of all nodes in the bookstore platform |V|×d , P({v m-(k / 2) ,…,v m+(k / 2) }\v m |E(v m )) indicates that node v m The embedding vector predicts the random walk sequence except v m The likelihood probability function of the embedding vector of other nodes except m ) represents node v m The embedding vector, v m-(k / 2) represents the node with subscript m-(k / 2) in the random walk sequence of length k; S34: Based on the four types of user behaviors in the bookstore network: purchase, add to cart, click, and friendship between users, it is divided into four sub-views: "buy", "collect", "click", and "friend"; S35: Each user node u i The embedding representation vectors for specific user behaviors are generated independently in different behavior views; this process is expressed as: in, Represents user node u i When the user behavior type is behavior (i) The aggregate vector of the k-th order neighbors on the behavioral view of Represents user node u i The aggregation vector of the k-1th order neighbors, Indicates that the user behavior type is behavior (i) The set of neighbor nodes in the view of Represents the aggregation vector The neighbor node element v contained in j Belongs to user node u i The set of neighbor nodes mean(.) represents the mean aggregation function, and σ is the activation function; S36: User node u i The embedding vectors under each subview are concatenated into the overall behavior embedding vector; S37: Generate user attribute embedding vector u i(attr) ; The process is expressed as: in, Denotes the defined hyperparameters, D Z is the feature transformation matrix for different types of nodes, x i is node u i Initialization attribute feature vector, T represents the transpose operation; S38: Fuse the user neighborhood embedding vector, user behavior embedding vector, and user attribute embedding vector into the final overall embedding vector; this process is expressed as: in, Represents user node u i In user behavior type behav (i) The overall embedding vector u contains user neighborhood information, user behavior information, and user attribute information. i(neighbor) Represents user node u i The neighborhood embedding vector of Indicates that the behavior type is behav (i) The importance coefficient of user behavior embedding relative to the overall embedding, represents the feature transformation matrix, T represents the transpose operation, u i(behavior) Represents user node u i The overall behavior embedding vector, It is the user behavior type calculated based on the self-attention mechanism (i) Next, embed the vector u for the overall behavior i(behavior) The linear combination coefficients of W behav(i) and w beha(i) is a trainable matrix of different dimensions under user behavior type behav(i), D Z is the feature transformation matrix for different types of nodes, x i is node u i The initialization attribute feature vector of .
4. The method for aligning users of cross-e-commerce bookstores based on tensor decomposition and global representation according to claim 1 is characterized in that: Use the user interest group block index model to divide e-commerce bookstore users into interest groups. Divide e-commerce bookstore users into interest groups, including: S41: Construct an initial book category association matrix to quantify the degree of association between each book subcategory. The initial book association matrix is an all-zero matrix. S42: Calculate the correlation coefficient between the two book categories A and B, and compare it with the user u i The correlation coefficients between all related book subcategories are filled into the book category correlation matrix to obtain the final correlation matrix; S43: Extract each column in the final association matrix, aggregate the top k book subcategories with the highest association in each column into abstract book interest categories, and assign each user to each abstract book interest category to form a user interest group based on the user's actual purchase record.
5. The method for aligning users of cross-e-commerce bookstores based on tensor decomposition and holistic representation according to claim 1 is characterized in that: Calculate the correlation coefficient between two book subcategories A and B, expressed as: Among them, R (typA(i),ttpe(B)) represents the correlation coefficient between two book categories A and B, U Attention(ui,type(·)) Represents user u i The degree of interest in a certain subcategory of books. type(A) represents subcategory A, and type(B) represents subcategory B.
6. The method for aligning users of cross-e-commerce bookstores based on tensor decomposition and global representation according to claim 1, characterized in that: Calculate the similarity between the embedding vectors of user nodes in the same interest group, including: in, represents the cosine similarity of the embedding vectors of user 1 and user 2 on platforms x and y, represents the embedding vector of user 1 in the bookstore platform x, represents the embedding vector of user 2 in the bookstore platform y, and S1 represents the vector The modulus of S2 represents the vector The modulus length, s 1i Represents a vector The i-th element in s 2i Represents a vector The i-th element in n represents the vector or The number of elements in .