E-commerce platform user preference identification method based on missing visual angle semi-supervised classification

Abstract

An e-commerce platform user preference identification method based on missing view angle semi-supervised classification belongs to the technical field of e-commerce user preference identification, and comprises the following steps: constructing an original multi-view angle data set according to data of different view angles of a user, constructing a view angle specific graph matrix set based on the original multi-view angle data set, supplementing missing view data of the user in combination with a graph neural network; calculating instance-level contrast loss and category-level contrast loss, and performing back propagation to update a trainable parameter matrix of the graph neural network; connecting the multi-view-angle data to obtain fusion features of user view angles; constructing a global graph matrix based on a view angle specific graph matrix set, and carrying out semi-supervised classification on users whose purchase preferences are not identified according to fusion features of the global graph matrix and the user view angle; estimating prediction loss, and executing back propagation to update global graph features and trainable parameters; and repeatedly training the model, and outputting a user preference prediction label. According to the method, the user preference identification accuracy on the user data with the missing view angle can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce user preference identification, and in particular to an e-commerce platform user preference identification method based on missing perspective semi-supervised classification. Background Art

[0002] As e-commerce platforms continue to grow, the scale of goods sold has increased dramatically, and the content that users are interested in is often drowned in the massive amount of product information. Therefore, how to accurately identify users' purchasing preferences and then recommend products that meet their needs and preferences has become an urgent issue. E-commerce platforms usually actively ask users for shopping preference information when users register and initially log in to the platform, but due to user experience and privacy restrictions, the number of users willing to accept inquiries at this stage is limited.

[0003] Considering that a large amount of user behavior data is stored in the e-commerce platform system, which contains rich user information, including user purchase behavior, browsing history, social interaction, etc., most of the existing user preference identification methods are based on collaborative filtering to mine potential associations between users, which can effectively infer user purchase preference information through similar behaviors between users. However, these methods mainly rely on a single type of user behavior, such as browsing history, and extract a single relationship, which cannot capture multi-source complex relationships. Therefore, in order to make full use of the multi-source heterogeneous behavior information of users, many methods consider integrating the multi-source heterogeneous behavior information of users based on multi-view learning to fully explore the value of fusion information such as user purchase behavior, browsing history, and social interaction, and comprehensively characterize the user behavior preference characteristics. However, in the e-commerce platform system, the purchase behavior, browsing history, social interaction and other behavior data of some users do not exist at the same time. This may be due to system failure or the user has not purchased any goods. Running a user preference identification method based on multi-view learning on data with missing perspectives will introduce learning bias due to over-learning of available perspectives. Therefore, when the missing rate of perspectives is large, the accuracy of user preference identification is significantly reduced, and the user's shopping experience will also deteriorate.

[0004] Therefore, in order to solve the above problems, it is urgent to propose a method for identifying user preferences on e-commerce platforms based on missing perspective semi-supervised classification. Summary of the invention

[0005] The purpose of the present invention is to provide an e-commerce user preference identification method based on missing perspective semi-supervised classification to solve the problem of low accuracy in user preference identification on user data with missing perspective.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is a method for identifying user preferences of e-commerce platforms based on missing perspective semi-supervised classification, which includes the following steps:

[0007] S1. Construct the original multi-view dataset and indication matrix based on the data from different user perspectives, and construct the label matrix based on the user preference information;

[0008] S2, construct a set of view-specific graph matrices based on the original multi-view dataset;

[0009] S3, supplement the user's missing perspective data based on the perspective-specific graph matrix set combined with the graph neural network;

[0010] S4, calculate instance-level contrast loss and category-level contrast loss, and back-propagate to update the trainable parameter matrix of the graph neural network;

[0011] S5, connecting multi-view data to obtain the fusion features of the user's perspective;

[0012] S6, constructing a global graph matrix based on a set of view-specific graph matrices;

[0013] S7, semi-supervised classification of users whose purchase preferences are not identified based on the fusion features of the global graph matrix and user perspective;

[0014] S8, evaluate the prediction loss based on the cross entropy error of the user's purchase preference, and perform back propagation to update the global graph features and trainable parameters;

[0015] S9. Repeat the training of the model and output the user preference prediction label.

[0016] Furthermore, in step S1, the original multi-view dataset is represented as X={X (1) , X (2) , X (3)}, where X represents the original multi-view dataset, X (1) represents the purchase behavior data, X (2) Indicates browsing history data, X (3) Represents social interaction data;

[0017] The indicator matrix is ​​represented as W∈{0,1} n×3 , where W is the indicator matrix and n is the number of users;

[0018] The label matrix is ​​expressed as Where Y is the label matrix, is a set of real numbers, Ω is the set of users who have marked their purchase preferences, |Ω|<<n, and c is the number of product categories that users prefer to purchase.

[0019] Furthermore, in step S2, constructing a perspective-specific image matrix set includes the following steps:

[0020] S21. Calculate the cosine similarity between each pair of users in each perspective based on the user's perspective data to obtain a similarity matrix set of users in each perspective In the formula, F (v) is the similarity matrix of the user’s v-th perspective, is a set of real numbers, n is the number of users, represents the missing view indicator value of the vth view of the ith user;

[0021] S22. If the user lacks data of a certain perspective, search for the user's nearest neighbors in other perspectives, obtain the average of the perspective data corresponding to the missing perspective of the user's nearest neighbors, and supplement the user's perspective data;

[0022] S23. After supplementing the missing user perspective data, construct the K nearest neighbor graph matrix of all users in each perspective, and construct a perspective-specific graph matrix set S (v) Represents the image matrix of the v-th view.

[0023] Furthermore, in step S3, supplementing the missing user perspective data includes the following steps:

[0024] S31. Initialize the graph neural network trainable parameter matrix corresponding to the perspective and obtain the initial user feature embedding corresponding to the perspective Where U (v) represents the initial user feature embedding of the vth view, X (v) Represents the original data of the user's vth perspective, represents the trainable parameter matrix of the graph neural network at the vth perspective, and gnn represents the graph neural network;

[0025] S32. Obtain neighbor node information corresponding to the viewing angle:

[0026]

[0027] In the formula, is the neighbor node information with the aggregation hop count of Q from the vth perspective, Q ≥ 1, α is the restart probability, is the symmetric normalized adjacency matrix of the v-th view, is the neighbor node information with the aggregation hop count of Q-1 for the vth perspective, D (v) is the degree matrix, S (v) Represents the image matrix of the vth view, I n is the identity matrix;

[0028] S33. Obtain a multi-view data set corresponding to the view:

[0029]

[0030] In the formula, Represents the multi-view data of the vth view;

[0031] S34, supplement the missing perspective data of other perspectives, if Then Grant Otherwise Grant As shown below:

[0032]

[0033] In the formula, represents the complete multi-view data of the i-th user's v-th view, represents the multi-view data of the i-th user's v-th view, represents the original data of the i-th user's v-th perspective, represents the missing view indicator value of the vth view of the ith user;

[0034] Get initialized full multi-view data Indicates the completed user purchase behavior data. Indicates the completed user browsing history data. Represents the completed social interaction data.

[0035] Furthermore, in step S4, the instance-level contrast loss is expressed as:

[0036]

[0037] In the formula, is the instance-level contrast loss, O (v，u) is the similarity matrix between the embedding representations of the vth view and the uth view, diag is a diagonal matrix, I n is the identity matrix;

[0038] The class-level contrast loss is expressed as:

[0039]

[0040] In the formula, is the category-level contrast loss of the v-th view, Ω is the set of users with marked purchase preferences, and Φ i represents the set of users who have the same purchase preference as the i-th user. is the complete multi-view data of the i-th user’s v-th view, is the complete multi-view data of the j-th product at the v-th view. represents the complete multi-view data of the vth view of the tth user, log is the logarithmic function, sim(·) represents the cosine similarity calculation function, exp is the exponential function, and τ is a hyperparameter.

[0041] Furthermore, in step S4, the trainable parameter matrix of the updated graph neural network is:

[0042] Calculate instance-level-category-level dual contrast loss

[0043]

[0044] In the formula, is the instance-level contrast loss, is the category-level contrast loss of the v-th view;

[0045] According to the dual contrast loss Perform a back propagation to update and optimize the trainable parameter matrix set of the graph neural network Represents the trainable parameter matrix of the graph neural network for the vth view, and calculates the reconstruction loss for all views:

[0046]

[0047] In the formula is the reconstruction loss, n is the number of users, is the complete multi-view data of the i-th user’s v-th view, d (v) is the decoder of the vth view, is the embedding representation matrix of the i-th user’s v-th perspective, By the encoder e of the vth view (v) Calculated;

[0048] According to the reconstruction loss Perform a backpropagation pass to update the trainable parameters of the optimized autoencoder.

[0049] Furthermore, in step S6, constructing a global graph matrix based on a set of view-specific graph matrices includes the following steps:

[0050] S61. Construct a weighted adjacency matrix of a set of view-specific graph matrices:

[0051]

[0052] Where S is the weighted adjacency matrix, π (v) is the specific graph weight coefficient of the vth view, D (v) is the degree matrix of the vth view, S (v) Represents the image matrix of the vth view, I nis the identity matrix, st represents the constraint condition;

[0053] S62. Constructing a learnable global graph:

[0054] ρ(S)=S⊙ReLU(Sigmoid(1 / 2(Γ+Γ T ))-M)

[0055] Where ρ(S) is the global graph, ⊙ is the Hadamard product, ReLU is the rectified linear activation function, Sigmoid is the logistic activation function, Γ is the learnable matrix, Γ T is the transposed matrix of Γ, M is the threshold matrix,

[0056] The global graph matrix is ​​denoted as A = ρ(S).

[0057] Further, in step S7, the semi-supervised classification comprises the following steps:

[0058] S71. Using the fusion features of the global graph matrix A and the user's perspective Calculate the classification probability matrix and obtain the user's purchase preference classification probability matrix, which is:

[0059]

[0060] Where Z is the user's purchase preference classification probability matrix, softmax is the soft maximum activation function, σ(·) is the activation function, is the trainable parameter matrix of the first layer graph convolutional network, is the trainable parameter matrix of the second-layer graph convolutional network, Indicates the completed user purchase behavior data. Indicates the completed user browsing history data. Represents the completed social interaction data.

[0061] S72. Obtaining predicted purchase preferences of all users argmax is the maximum parameter function.

[0062] Furthermore, in step S8, the predicted loss is:

[0063]

[0064] In the formula, To predict the loss, Y i,j represents the true preference of the i-th user for the j-th product, Y i,j ∈[0,1], ln is the logarithmic function, Z i,j represents the probability that the ith user prefers the jth product, c is the number of product categories that the user prefers to purchase.

[0065] Furthermore, step S9 further includes:

[0066] The model is trained until the variance of the prediction loss in the model training iteration is less than the threshold, which is set to 0.01.

[0067] The beneficial effects of the present invention are:

[0068] (1) Based on the similarity relationship between the behavioral data of all users, the present invention constructs perspective-specific graphs in three different behavioral perspectives: purchase behavior, browsing history, and social interaction, and learns a refined global graph based on the perspective-specific graphs. The perspective-specific graph provides conditions for mining perspective-specific information, and the global graph integrates complementary information from multiple perspectives.

[0069] (2) The present invention uses instance-level and category-level contrastive learning combined with perspective-specific graphs to complete the user's missing behavioral perspective. Instance-level contrastive learning uses cross-perspective consistency to guide the completion process, while category-level contrastive learning applied within a perspective uses the preference information of some users collected by the previous e-commerce platform system (i.e., a small amount of supervisory signals) to enhance the category discrimination ability of the completed perspective and promote the learning of perspective-specific information.

[0070] (3) Based on the global graph, the present invention applies GCN and a small amount of supervisory signals to perform semi-supervised classification of users to identify their purchasing preferences, which not only avoids decision divergence between multiple perspectives, but also effectively propagates user preference information in non-Euclidean space. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0072] Figure 1 It is a schematic diagram of the framework of the present invention.

[0073] Figure 2 It is a schematic diagram of the process of the present invention.

[0074] Figure 3 It is a schematic diagram of an incomplete multi-view data set according to Example 1 of the present invention.

[0075] Figure 4 These are the perspective-specific graph, the weighted fusion graph and the global graph of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0076] 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.

[0077] like Figure 1 , Figure 2 As shown, the present invention discloses an e-commerce user preference identification method based on missing perspective semi-supervised classification, comprising the following steps:

[0078] S1. Collect data from three perspectives: user purchase behavior, browsing history, and social interaction, construct the original multi-perspective dataset X and the indicator matrix W that describes the distribution of missing perspectives, and collect user preference information based on the e-commerce platform system to form a label matrix Y. The specific process is as follows:

[0079] Collect user purchase behavior data X(1) from the order system of the e-commerce platform system. Where n is the number of users, m is the number of product categories, is a real number set, and the purchase behavior data of the i-th user purchasing the j-th product is expressed as The corresponding data of users who have not purchased any products are all set to 0.

[0080] Collect user browsing history data from the user system of the e-commerce platform system X (2) , The browsing record data of the i-th user browsing the j-th product is expressed as The corresponding data of users who have not browsed the products are all set to 0.

[0081] Collect all users' social interaction data from the social system of the e-commerce platform system X (3) , m 3 is the number of user social circles, and the interaction of the i-th user in the k-th social circle is expressed as The corresponding data of users who have not participated in social circle interactions are all set to 0.

[0082] User purchase behavior data X (1) , Browsing history dataX (2) and social interaction dataX (3) After preprocessing, the original multi-view dataset X = {X (1) , X (2) , X (3)} and indicator matrix W, W∈{0,1}n×3 , if the vth perspective of the ith user is missing, it is represented as otherwise Where v is used to identify different perspectives, v=1 represents the purchase behavior perspective, v=2 represents the browsing history perspective, and v=3 represents the social interaction perspective.

[0083] The label matrix Y is constructed based on the user purchase preference information collected by the e-commerce platform system. Ω is the set of users who have marked their purchase preferences, |Ω|<<n, and c is the number of product categories that users prefer to purchase.

[0084] S2. Construct a view-specific graph matrix set based on the original multi-view dataset, including the following steps:

[0085] S21. Calculate the cosine similarity between each pair of users in each perspective based on the user's perspective data to obtain a similarity matrix of users in each perspective In the formula, F (v) is the similarity matrix of the user’s v-th perspective, Represents the missing view indicator value of the vth view of the ith user.

[0086] S22. If the user is missing data of a certain perspective, search for the user's nearest neighbors in other perspectives, obtain the average value of the corresponding perspective data of the user's nearest neighbors in the missing perspective, and fill in the user's perspective data.

[0087] S23. After supplementing the missing user perspective data, the KNN algorithm (K nearest neighbor algorithm) is used to construct the K nearest neighbor graph matrix of all users in each perspective, and a perspective-specific graph matrix set is constructed. S (v) represents the v-th view-specific image matrix.

[0088] S3, supplementing the missing user perspective data based on the perspective-specific graph matrix set combined with the graph neural network, including the following steps:

[0089] S31. Initialize the graph neural network trainable parameter matrix corresponding to the perspective and obtain the initial user feature embedding corresponding to the perspective Where U (v) represents the initial user feature embedding of the vth view, X (v) Represents the original data of the user's vth perspective, represents the trainable parameter matrix of the graph neural network at the vth perspective, and gnn represents the graph neural network;

[0090] X (v) The missing user data in is filled with zeros, and set Indicates the neighbor node information with aggregation hop count of 0 at the v-th perspective.

[0091] S32. Obtain neighbor node information corresponding to the viewing angle:

[0092]

[0093]

[0094] In the formula, is the neighbor node information with the aggregation hop count of Q from the vth perspective, Q ≥ 1, α is the restart probability, α∈(0, 1], is the symmetric normalized adjacency matrix with self-loops of the v-th view, is the neighbor node information with the aggregation hop count of Q-1 for the vth perspective, D (v) is the degree matrix, S (v) Represents the image matrix of the vth view, I n is the identity matrix;

[0095] S33. Obtain a multi-view data set corresponding to the view:

[0096]

[0097] In the formula, Represents the multi-view data of the vth view;

[0098] S34, supplement the missing perspective data of other perspectives, if Then Grant Otherwise Grant As shown below:

[0099]

[0100] In the formula, represents the complete multi-view data of the i-th user's v-th view, represents the multi-view data of the i-th user's v-th view, represents the original data of the i-th user's v-th perspective, Represents the missing view indicator value of the vth view of the ith user.

[0101] Get initialized full multi-view data Indicates the completed user purchase behavior data. Indicates the completed user browsing history data. Represents the completed social interaction data,

[0102] S4. Calculate the instance-level contrast loss (ICL) and category-level contrast loss (CCL), and back-propagate to update the trainable parameter matrix of the graph neural network, which is:

[0103] An autoencoder (AE) is introduced in each behavior view, and the autoencoder in the vth view is denoted as AE (v) , by the encoder and decoder Composition, e (v) is the encoder of the vth view, d (v) is the decoder of the vth view, For encoder e (v) The trainable parameters of the v-th view, For decoder d (v) The trainable parameters of the v-th view are and Initialize.

[0104] Get the embedding representation matrix set H of the user from each perspective, which is:

[0105]

[0106] In the formula, H (v) is the embedding representation matrix of the vth view, and b is the dimension of the view embedding representation.

[0107] Calculate the difference between the i-th user in the v-th perspective and the p-th user in the u-th behavior perspective for:

[0108]

[0109] In the formula, is the embedding representation of the corresponding data of the i-th user in the v-th view, which is obtained by the encoder e (v) go through Calculated, is the embedding representation of the corresponding data of the pth user in the uth behavior perspective, and T is the transposed matrix;

[0110] Calculate the similarity matrix O between the embedding representations of the vth view and the uth action view (v，u) , To prevent users belonging to the same purchase preference category from being stretched, the present invention will (v，u) All elements on the diagonal are constrained to be 1, and no constraints are imposed on other elements. The instance-level contrast loss is calculated:

[0111]

[0112] In the formula, represents the instance-level contrast loss, In represents the identity matrix, diag is the diagonal matrix, I n is the identity matrix;

[0113] Calculate the category-level contrast loss as:

[0114]

[0115] In the formula, represents the category-level contrast loss of the vth view, Ω is the user who has marked the purchase preference information, and Φ i represents the set of users who have the same purchase preference as the i-th user. is the complete multi-view data of the j-th product at the v-th view. represents the complete multi-view data of the vth view of the tth user, log is the logarithmic function, sim(·) represents the cosine similarity calculation function, exp is the exponential function, and τ is a hyperparameter, which is used to control the distribution of similarity scores and thus adjust the model's attention to difficult samples.

[0116] Calculate the instance-level-category-level dual contrast loss:

[0117]

[0118] In the formula, is the instance-level-category-level dual contrast loss,

[0119] According to the dual contrast loss Perform a back propagation to update and optimize the trainable parameter matrix set of the GNN completion network Calculate the reconstruction loss for all views

[0120]

[0121] According to the reconstruction loss Perform a backpropagation to update and optimize the trainable parameters of the autoencoder

[0122] S5. Connecting multi-view data Obtaining fusion features of user perspective

[0123] S6. Since the generation of view-specific graphs may lead to inaccurate user node similarity relationships, directly combining and weighting the view-specific graphs will amplify this adverse effect. In order to integrate complementary information from multiple viewpoints, the present invention constructs a global graph matrix based on a set of view-specific graph matrices, including the following steps:

[0124] S61. Construct a weighted adjacency matrix of a set of view-specific graph matrices:

[0125]

[0126] Where S is the weighted adjacency matrix, π (v) is the specific graph weight coefficient of the vth view, D (v) is the degree matrix of the vth view, and st represents the constraint condition;

[0127] S62: Building a learnable global graph:

[0128] ρ(S)=S⊙ReLU(Sigmoid(1 / 2(Γ+Γ T ))-M)

[0129] Where ρ(S) is the global graph, ⊙ is the Hadamard product, ReLU is the rectified linear activation function, Sigmoid is the logistic activation function, Γ is the learnable matrix, It is possible to learn edge-specific coefficients for undirected graphs, Γ T is the transposed matrix of Γ, M is the threshold matrix, Used to control node relationship activation;

[0130] Then the global graph matrix is ​​expressed as A=ρ(S).

[0131] The global graph matrix A can eliminate low-quality edges in the graph structure based on the weighted adjacency matrix S. In the present invention, low-quality edges indicate that there is a low similarity between two users.

[0132] S7. Based on the fusion features of the global graph matrix and the user perspective, a GCN classifier containing two layers of GCN is used to perform semi-supervised classification on users whose purchase preferences are not identified. The following steps are included:

[0133] S71. Using the fusion features of the global graph matrix A and the user's perspective Input the GCN classifier, calculate the classification probability matrix, and obtain the user's purchase preference classification probability matrix Z, which is expressed as:

[0134]

[0135] In the formula, softmax is the soft maximum activation function, σ(·) is the activation function, is the trainable parameter matrix of the first layer graph convolutional network, is the trainable parameter matrix of the graph convolutional network of the second layer;

[0136] S72. Obtaining predicted purchase preferences of all users argmax is the maximum parameter function;

[0137] The GCN network can integrate connection patterns and feature attributes with given graph structure data, propagate node features to adjacent networks, and make full use of existing node and label information.

[0138] S8. Evaluate the prediction loss of the GCN classifier based on the cross entropy error of the user's purchase preference, as shown in the following formula:

[0139]

[0140] In the formula, To predict the loss, Y i,j represents the true preference of the i-th user for the j-th product, Y i,j ∈[0,1],Y i,j =1 means user i prefers product j, Y i,j =0 means no preference, ln is the logarithmic function, Z i,j represents the probability that the ith user prefers the jth product,

[0141] Perform backpropagation to update the global graph and trainable parameters in the GCN classifier, including the following steps:

[0142] S81. Optimize the trainable parameter matrix of the GNN completion network

[0143] S82. Optimizing the trainable parameters of the autoencoder and

[0144] S83. Optimizing the global graph and trainable parameters in the GCN classifier Γ, M, and

[0145] In a separate training iteration, each step performs a forward pass, followed by a backward pass with the uncorrelated variables fixed.

[0146] S9. Repeat the model training until the prediction loss variance in the model training iteration is less than the threshold value of 0.01, and output the user preference prediction label

[0147] Example 1

[0148] According to the present invention, a method for identifying e-commerce user preferences based on missing view semi-supervised classification is used to identify user preferences. The schematic diagram of the incomplete multi-view dataset is shown in Figure 3 As shown, the subscripts in the figure are used to distinguish different users, and the superscripts in brackets are used to distinguish different behavioral perspectives, e.g. Indicates the purchase behavior perspective of View1 of user 1. (?) indicates that the user is missing in the perspective (i.e., the system has not recorded such behavior of the user) or the user is missing the purchase preference category (i.e., the system has not yet identified the user's purchase preference). The number of neighbors of each node defaults to 2, and the predicted preferences of all users Among them, 0 means the preference is for trendy clothing, and 1 means the preference is for computer DIY.

[0149] like Figure 3 As shown in the figure, according to F(2) and F(3), we can find the missing perspective data of users in the purchasing behavior perspective of View1. The nearest neighbors in View2 browsing history and View3 social interaction are and use Fill in the behavior data corresponding to missing users use Fill in the behavior data corresponding to the missing users in the View2 browsing history perspective use Fill in the missing user behavior data in View3 social interaction perspective

[0150] by Figure 3 As an example of the incomplete multi-view dataset in The fusion features of all user behavior perspectives are obtained as follows:

[0151]

[0152] by Figure 3 Taking the simplified incomplete multi-view dataset as an example, according to and right Assign values ​​as follows:

[0153] In View1

[0154] In View2

[0155] In View3

[0156] Figure 4 The view-specific graph, weighted fusion graph, and global graph of Example 1. 1 ,according to Figure 4 The global graph in the figure is composed of nodes X from its neighbors. 3 , X 4 The discriminant information of 1 With X 3Closer, so X 1 The final predicted purchase preference of 1 More inclined to y3 (computer DIY).

[0157] by Figure 3 Taking the incomplete multi-view dataset in as an example, we obtain the neighbor node information of each view with an aggregation hop count of 2 (since there are only four user nodes, the distance from a randomly selected node to all nodes in the graph structure will not exceed 2) in For example, in user X 2 In the View1 purchasing behavior perspective according to Figure 4 In the view-specific graph of View1, we first aggregate the 1-hop distance neighbors Based on this information, the 2-hop distance neighbors are aggregated again Therefore, after the two-hop neighbor node information aggregation, the missing perspective Completed and enriched with specific information from the View1 purchasing behavior perspective.

[0158] Example 2

[0159] To verify the effectiveness of the present invention, a comparative experiment on classification accuracy (Acc) was conducted on the ALOI, BBCnews, BBCsports, and Caltech101 datasets with the existing incomplete multi-view semi-supervised classification methods SLIM, AMSC, and DIMC. The experimental results are shown in Table 1. PER represents the sample ratio of missing viewpoints, where ' / ' indicates that the model reports an error due to timeout or insufficient memory during operation. The comparative experiment was conducted with a labeled sample ratio of 10%.

[0160] Table 1 Comparative experiments with the most advanced incomplete multi-view semi-supervised classification methods

[0161]

[0162]

[0163] Experimental results show that the proposed method has excellent classification accuracy under different PREs, and provides a solution to the problem of low accuracy in user preference recognition in the missing perspective.

[0164] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0165] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for identifying user preferences on an e-commerce platform based on missing perspective semi-supervised classification, characterized in that: The following steps are involved: S1. Construct an original multi-view dataset and an indication matrix based on the data from different user perspectives, and construct a label matrix based on the existing user preference information; S2, construct a set of view-specific graph matrices based on the original multi-view dataset; S3, supplement the user's missing perspective data based on the perspective-specific graph matrix set combined with the graph neural network; S4, calculate instance-level contrast loss and category-level contrast loss, and back-propagate to update the trainable parameter matrix of the graph neural network; S5, connecting multi-view data to obtain the fusion features of the user's perspective; S6, constructing a global graph matrix based on a set of view-specific graph matrices; S7, semi-supervised classification of users whose purchase preferences are not identified based on the fusion features of the global graph matrix and user perspective; S8, evaluate the prediction loss based on the cross entropy error of the user's purchase preference, and perform back propagation to update the global graph features and trainable parameters; S9. Repeat the training of the model and output the user preference prediction label.

2. According to claim 1, a method for identifying user preferences of e-commerce platforms based on missing perspective semi-supervised classification is characterized in that: In step S1, the original multi-view data set is represented as X={X (1) ,X (2) ,X (3) }, where X represents the original multi-view dataset, X (1) represents the purchase behavior data, X (2) Indicates browsing history data, X (3) Represents social interaction data; The indicator matrix is ​​represented as W∈{0,1} n×3 , where W is the indicator matrix and n is the number of users; The label matrix is ​​expressed as Where Y is the label matrix, is a set of real numbers, Ω is the set of users who have marked their purchase preferences, |Ω|<<n, and c is the number of product categories that users prefer to purchase.

3. According to claim 1, the method for identifying user preferences of e-commerce platforms based on missing perspective semi-supervised classification is characterized in that: In step S2, constructing a view-specific image matrix set includes the following steps: S21. Calculate the cosine similarity between each pair of users in each perspective based on the user's perspective data to obtain a similarity matrix set of users in each perspective In the formula, F (v) is the similarity matrix of the user’s v-th perspective, is a set of real numbers, n is the number of users, represents the missing view indicator value of the vth view of the ith user; S22. If the user lacks data of a certain perspective, search for the user's nearest neighbors in other perspectives, obtain the average of the perspective data corresponding to the missing perspective of the user's nearest neighbors, and supplement the user's perspective data; S23. After supplementing the missing user perspective data, construct the K nearest neighbor graph matrix of all users in each perspective, and construct a perspective-specific graph matrix set S (v) Represents the image matrix of the v-th view.

4. According to claim 1, the method for identifying user preferences of e-commerce platforms based on missing perspective semi-supervised classification is characterized in that: In step S3, supplementing the missing viewing angle data of the user comprises the following steps: S31. Initialize the graph neural network trainable parameter matrix corresponding to the perspective and obtain the initial user feature embedding corresponding to the perspective Where U (v) represents the initial user feature embedding of the vth view, X (v) Represents the original data of the user's vth perspective, represents the trainable parameter matrix of the graph neural network at the vth perspective, and gnn represents the graph neural network; S32. Obtain neighbor node information corresponding to the viewing angle: In the formula, is the neighbor node information with the aggregation hop count of Q from the vth perspective, Q ≥ 1, α is the restart probability, α∈(0,1], is the symmetric normalized adjacency matrix of the v-th view, is the neighbor node information with the aggregation hop count of Q-1 for the vth perspective, D (v) is the degree matrix, S (v) Represents the image matrix of the vth view, I n is the identity matrix; S33. Obtain a multi-view data set corresponding to the view: In the formula, Represents the multi-view data of the vth view; S34, supplement the missing perspective data of other perspectives, if Then Grant Otherwise Grant As shown below: In the formula, represents the complete multi-view data of the i-th user's v-th view, represents the multi-view data of the i-th user's v-th view, represents the original data of the i-th user's v-th perspective, represents the missing view indicator value of the vth view of the ith user; Get initialized full multi-view data Indicates the completed user purchase behavior data. Indicates the completed user browsing history data. Represents the completed social interaction data.

5. According to claim 1, the method for identifying user preferences of e-commerce platforms based on missing perspective semi-supervised classification is characterized in that: In step S4, the instance-level contrast loss is expressed as: In the formula, is the instance-level contrast loss, O (v,u) is the similarity matrix between the embedding representations of the vth view and the uth view, diag is a diagonal matrix, I n is the identity matrix; The class-level contrast loss is expressed as: In the formula, is the category-level contrast loss of the vth view, Ω is the set of users with marked purchase preferences, and Φ i represents the set of users who have the same purchase preference as the i-th user. is the complete multi-view data of the i-th user’s v-th view, is the complete multi-view data of the j-th product at the v-th view, represents the complete multi-view data of the vth view of the tth user, log is the logarithmic function, sim(·) represents the cosine similarity calculation function, exp is the exponential function, and τ is a hyperparameter.

6. The method for identifying user preferences of an e-commerce platform based on missing perspective semi-supervised classification according to claim 1, characterized in that: In step S4, the trainable parameter matrix of the updated graph neural network is: Compute instance-level-category-level dual contrast loss In the formula, is the instance-level contrast loss, is the category-level contrast loss of the v-th view; According to the dual contrast loss Perform a back propagation to update and optimize the trainable parameter matrix set of the graph neural network Represents the trainable parameter matrix of the graph neural network for the vth view, and calculates the reconstruction loss for all views: In the formula is the reconstruction loss, n is the number of users, is the complete multi-view data of the i-th user’s v-th view, d (v) is the decoder of the vth view, is the embedding representation matrix of the i-th user’s v-th perspective, By the encoder e of the vth view (v) Calculated; According to the reconstruction loss Perform a backpropagation pass to update the trainable parameters of the optimized autoencoder.

7. According to claim 1, the method for identifying user preferences of e-commerce platforms based on missing perspective semi-supervised classification is characterized in that: In step S6, constructing a global graph matrix based on a set of view-specific graph matrices includes the following steps: S61. Construct a weighted adjacency matrix of a set of view-specific graph matrices: Where S is the weighted adjacency matrix, π (v) is the specific graph weight coefficient of the vth view, D (v) is the degree matrix of the vth view, S (v) Represents the image matrix of the vth view, I n is the identity matrix, st represents the constraint condition; S62. Constructing a learnable global graph: ρ(S)=S⊙ReLU(Sigmoid(1 / 2(Γ+Γ) T ))-M) Where ρ(S) is the global graph, ⊙ is the Hadamard product, ReLU is the rectified linear activation function, Sigmoid is the logistic activation function, Γ is the learnable matrix, Γ T is the transposed matrix of Γ, M is the threshold matrix, The global graph matrix is ​​denoted as A = ρ(S).

8. The method for identifying user preferences of an e-commerce platform based on missing perspective semi-supervised classification according to claim 1, characterized in that: In step S7, the semi-supervised classification comprises the following steps: S71. Using the fusion features of the global graph matrix A and the user's perspective Calculate the classification probability matrix and obtain the user's purchase preference classification probability matrix, which is: Where Z is the user's purchase preference classification probability matrix, softmax is the soft maximum activation function, σ(·) is the activation function, is the trainable parameter matrix of the first layer graph convolutional network, is the trainable parameter matrix of the second-layer graph convolutional network; Indicates the completed user purchase behavior data. Indicates the completed user browsing history data. Represents the completed social interaction data; S72. Obtaining predicted purchase preferences of all users argmax is the maximum parameter function.

9. The method for identifying user preferences of an e-commerce platform based on missing perspective semi-supervised classification according to claim 1, characterized in that: In step S8, the predicted loss is: In the formula, To predict the loss, Y i,j represents the true preference of the i-th user for the j-th product, Y i,j ∈[0,1], ln is the logarithmic function, Z i,j represents the probability that the ith user prefers the jth product, c is the number of product categories that the user prefers to purchase.

10. The method for identifying user preferences of an e-commerce platform based on missing perspective semi-supervised classification according to claim 1, characterized in that: Step S9 also includes: The model is trained until the variance of the prediction loss in the model training iteration is less than the threshold, which is set to 0.01.

Images