E-commerce network community discovery method and system based on semi-supervised network representation learning

Abstract

The present invention relates to the field of e-commerce network data mining and discloses a method and system for discovering e-commerce network communities based on semi-supervised network representation learning. The method comprises steps of constructing an e-commerce network adjacency matrix, a feature matrix, and a partial user label matrix; constructing and optimizing an e-commerce network community discovery model based on the semi-supervised network representation learning; and outputting the results of the e-commerce network community discovery model based on the semi-supervised network representation learning. The e-commerce network community discovery model based on the semi-supervised network representation learning is constructed using two independent GCN encoders, and a joint optimization objective function is designed, and the model is trained by minimizing the objective function. The system comprises a computer processor and memory, an e-commerce network adjacency matrix, a feature matrix, and a label matrix unit, a training unit for the e-commerce network community discovery model based on the semi-supervised network representation learning, and a unit for outputting the results of the e-commerce network community discovery model based on the semi-supervised network representation learning.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce network data mining, and in particular to an e-commerce network community discovery method and system based on semi-supervised network representation learning. Background Art

[0002] In today's society, with the rapid development of the internet, numerous large-scale e-commerce platforms have emerged, such as Taobao, JD.com, and Pinduoduo. These platforms are incredibly popular because they allow people to purchase what they need anytime, anywhere, bringing significant convenience to people's lives. However, when shopping on e-commerce platforms, people often encounter difficulties, such as difficulty finding desired items, discovering interesting activities, or finding recommendations and sharing experiences with people who share similar interests. To address these issues, researchers have studied the relationships between users, products, and stores in e-commerce networks. They found that users with the same or similar interests often share similar information, such as product browsing history, purchase history, product favorites, favorite stores, and events attended. This allows them to form communities. For users in the same community, platforms can recommend related products, stores, and events based on the interests of the entire community. These users can also engage in shared interests and product sharing. Researching communities within e-commerce networks can help accurately push products to platforms, improve platform purchase rates, and enhance user satisfaction. To this end, this paper proposes a method and system for discovering e-commerce network communities based on semi-supervised network representation learning. By learning user feature information, partial label information, and spatial structure information within the e-commerce network, an optimal low-dimensional representation of users is obtained. Then, using the k-means algorithm, users with similar interests are grouped into the same community, enabling better performance in tasks such as product recommendations, store recommendations, friend or group chat recommendations, user interest analysis, and interest community analysis. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to design an e-commerce network community discovery method and system based on semi-supervised network representation learning, which can effectively and reliably divide users in the e-commerce network.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] The present invention provides an e-commerce network community discovery method based on semi-supervised network representation learning. The method includes three main steps: step S10 is the construction step of the adjacency matrix, feature matrix and partial user label matrix of the e-commerce network; steps S20 to S30 are the construction and optimization steps of the e-commerce network community discovery model based on semi-supervised network representation learning; and step S40 is the result output step of the e-commerce network community discovery model based on semi-supervised network representation learning.

[0006] S10. Obtain an e-commerce network user dataset, construct an e-commerce network structure based on the friendship relationships between users, use users' purchase data, browsing data, store follow data, event participation data, and product collection data as user node features, use users with labeled types as label data, and construct an adjacency matrix, a feature matrix, and a label matrix based on the e-commerce network structure, user node features, and part of the label data;

[0007] S20, based on the adjacency matrix, feature matrix, and label matrix established in step S10, using two independent GCN encoders to build an e-commerce network community discovery model with semi-supervised network representation learning;

[0008] S30. For the e-commerce network community discovery model based on the semi-supervised network representation learning constructed in step S20, a joint optimization objective function is designed, and the model parameters are determined by training the model by minimizing the objective function.

[0009] S40 , using the e-commerce network community discovery model constructed by the semi-supervised network representation learning in step S20 and the model parameters determined in step S30 , to achieve community discovery and detection of the e-commerce network, and output the community discovery results.

[0010] Furthermore, step S10 of the method includes the following specific steps:

[0011] S11. Obtain and load the e-commerce network user dataset from the e-commerce network platform, build the e-commerce network structure based on the friendship relationship between users, use the user's purchase data, browsing data, store attention data and product collection data as user node features, and use the marked type of users as label data; denote the e-commerce network as G = (V, E, X), V = {v1, v2, ..., v N} represents the set of N user nodes in the e-commerce network, where the nth user (1 n N) is recorded as user node v n ; Represents a set of L user nodes of marked types in the e-commerce network, where the lth user (1 lL) is recorded as a user node of marked type E={e1,e2,…,e M} represents the M edges between user nodes, where the mth edge (1 m M) is denoted as e m ; X is an N×R dimensional user node feature matrix, whose nth (1 n N) row x n =[x n1 ,x n2 ,… x nR ] represents the user node v in the e-commerce network n R features of which element x nr Represents user node v n The rth feature of ; A represents an N×N adjacency matrix, corresponding to the M edges contained in the set E, and the element in the i-th row and j-th column (1 i N, 1j N) in the matrix is ​​A ij Represents the i-th user node v i and the jth user node v j Adjacency relationship, if user node v i and user node v j If there is a connecting edge between them, then A ij =1, otherwise A ij =0; the user node category labeling situation in the e-commerce network is recorded through the L×K dimensional label matrix Y, where L represents the number of user nodes with category labels, K represents the number of label categories, and the element in the i-th row and j-th column (1 i L, 1 j K) in the matrix is ​​Y ij Represents the user node of the i-th marked type The relationship between the j-th class label, if the user node of the marked type If it belongs to this category, then Y ij =1, otherwise Y ij =0.

[0012] Furthermore, the e-commerce network community discovery model of the semi-supervised network representation learning in step S20 includes three parts: a semi-supervised channel, a self-supervised channel, and a redundancy elimination module, wherein: the semi-supervised channel includes a GCN encoder, which uses the user information of the labeled type as a guide to learn the common features between users with the same interests and the difference features between users with different interests; the self-supervised channel also includes a GCN encoder, which constructs false user feature information and compares it with the user's real feature information, so that the user's low-dimensional representation can better capture and maintain the structure and characteristics of the e-commerce network under the real-false feature information comparison mechanism; the redundancy elimination module is used to remove the overlapping information in the above two channels, enhance the effectiveness of the guidance and supervision role in the user's low-dimensional representation learning process, so that the e-commerce network community discovery model obtains stable performance. Its overall structure is shown in the attached figure. Figure 1 As shown in the figure, the specific operation steps of the model are:

[0013] S21. In the semi-supervised channel, the e-commerce network G = (V, E, X) is input into the GCN encoder for encoding to obtain the representation vector of the user node in the e-commerce network. The formal representation of the encoding process is shown in formula (1):

[0014]

[0015] in, D is the degree matrix of the user node, user node v i The degree is expressed as the number of edges associated with the user node; W0 and W1 are learnable parameter matrices; H = {h1,h2…,h N} is the set of representation vectors of user nodes in the e-commerce network, h i Represents user node v i The low-dimensional representation of the user is obtained after encoding; the input here consists of two parts: the e-commerce network user node feature matrix X and the e-commerce network adjacency matrix A; the activation function Relu(·) in the encoder is defined by formula (2):

[0016]

[0017] When constructing the semi-supervised channel loss, the present invention uses a neural network classifier based on Softmax and cross entropy loss functions as a mutual information estimator to evaluate the correlation between the user node feature X and the partial user node label Y, so that the user low-dimensional representation H = {h1,h2…,h N} can learn more relevant information about the user node label Y; the semi-supervised channel optimization objective is shown in formula (3):

[0018]

[0019] Among them, x l Represents the features of the lth user node with label, y l represents the label value of the lth user node with a label, y′ represents the user label category, P(y′) is the prior distribution probability of the user label, f(x l ) represents the low-dimensional representation of the lth labeled user node obtained by the GCN encoder;

[0020] S22. In the self-supervisory channel, the e-commerce network user node feature matrix X and the e-commerce network adjacency matrix A are also used as input, and the user low-dimensional representation is obtained through the GCN encoder. Represents user node v i The low-dimensional representation of the user is obtained after encoding; when constructing the self-supervised channel loss, we use the Jensen-Shannon estimator to measure the low-dimensional representation of the user With user node vn The correlation between the user nodes and their adjacent nodes is measured from two aspects: the user's node feature X and the adjacency matrix A. The overall optimization goal of this channel is shown in formula (4):

[0021]

[0022] Among them, G i Indicates that the user node v i and its adjacent user nodes constitute the subgraph, Indicates the consistency of reconstruction of user node features, I(a ij ;w ij ) indicates the consistency of user structure reconstruction, Low-dimensional representation of users in e-commerce networks and The correlation, a ij Represents user node v i and v j The weight of the edge between them;

[0023] The consistency of user node feature reconstruction is calculated by formula (5):

[0024]

[0025] Among them, x j Represents user node v n And the feature information of the adjacent user nodes, x′ j It is a false user node feature generated by disrupting the adjacency matrix A of the user node, while keeping the feature matrix X unchanged, to generate the corresponding false user node feature information; is a discriminator used to discriminate between real user node features and generated false user node features, as shown in formula (6):

[0026]

[0027] Among them, Θ1 is a trainable score matrix, Where T represents the transposition operation and σ is the sigmoid function, as shown in formula (7):

[0028]

[0029] The consistency of the reconstruction of the user structure of the e-commerce network is calculated by (8):

[0030] I(a ij ;w ij )=a ij logw ij +(1-a ij )log(1-w ij) (8)

[0031] S23, using the redundancy elimination module to identify and measure redundant information, the user low-dimensional representation obtained in the semi-supervised channel and the self-supervised channel is h l and The correlation between the two is used to measure information redundancy, as shown in formula (9):

[0032]

[0033] in, It is aimed at Generated false data, is the discriminator, h l T Where T is the transposition operation, as shown in formula (10):

[0034]

[0035] Furthermore, step S30 of the method includes the following specific steps:

[0036] S31, the semi-supervised channel, self-supervised channel and redundancy elimination module described in S20 are jointly trained to define the overall optimization objective function for the e-commerce network community discovery model based on semi-supervised network representation learning as shown in formula (11):

[0037]

[0038] in, is the semi-supervised channel loss, is the self-supervised channel loss, I r is redundant information, θ is a hyperparameter used to measure the importance of the self-supervisory channel;

[0039] S32. Initialize the parameters of the e-commerce network community discovery model using the semi-supervised network representation learning in step S20. Given an adjustment coefficient θ and a number of iterations T, iteratively execute steps S33 to S35 until the number of iterations is reached, thereby completing the training of the e-commerce network community discovery model using the semi-supervised network representation learning and obtaining the optimal parameters of the model.

[0040] S33, the e-commerce network user adjacency matrix A and feature matrix X obtained in S11 are input into the semi-supervised channel and the self-supervised channel respectively, and the semi-supervised e-commerce network representation H = {h1,h2…,h N} and self-supervised e-commerce network representation

[0041] S34, according to formula (3), formula (4) and formula (9) respectively get the semi-supervised channel target loss Self-supervised channel target loss and redundant information I r ;

[0042] S35. Input the loss obtained in S34 into formula (11) to obtain the total loss of the e-commerce network community discovery model of the semi-supervised network representation learning, perform backpropagation, and cyclically update the semi-supervised channel neural network parameters W0 and W1 and the self-supervised channel neural network parameters And the score matrices Θ1, Θ2.

[0043] Furthermore, step S40 of the method includes the following specific steps:

[0044] S41, obtain the optimal parameters of the model through S33 to S35, input new e-commerce network user data, and obtain the community label of each user node through formula (12);

[0045]

[0046] Among them, c (i) represents the community label of the i-th user node, z (i) represents the representation of the i-th user node, z (i) Calculated by formula (13),

[0047]

[0048] μ (j) represents the centroid of the jth cluster, μ (j) Calculated by formula (14),

[0049]

[0050] S42. Output the e-commerce network community discovery results to e-commerce platform operators or researchers in related fields for tasks such as product recommendations, store recommendations, friend or group chat recommendations, user interest analysis, and interest community analysis.

[0051] The present invention also provides an e-commerce network community discovery system based on semi-supervised network representation learning, which is used to implement the above-mentioned e-commerce network community discovery method based on semi-supervised network representation learning, including a computer processor and memory, an e-commerce network adjacency matrix, a feature matrix and a label matrix unit, an e-commerce network community discovery model training unit based on semi-supervised network representation learning, and an e-commerce network community discovery model result output unit based on semi-supervised network representation learning; the e-commerce network adjacency matrix, feature matrix and label matrix unit execute step S10, for the acquired user data set, construct an e-commerce network structure according to the friendship relationship between users, use the user's purchase data, browsing data, store attention data, activity participation data and product collection data as user node features, and use the marked type of users as label data, thereby establishing the adjacency matrix, feature matrix and label matrix, and The method is loaded into the computer memory; the e-commerce network community discovery model training unit based on semi-supervised network representation learning executes steps S20 to S30 according to the adjacency matrix, feature matrix and label matrix constructed in step S10, constructs the e-commerce network community discovery model based on semi-supervised network representation learning, and trains the model based on the defined objective function; the e-commerce network community discovery model result output unit based on semi-supervised network representation learning executes step S40, divides users with similar interests in the e-commerce network into the same community, and outputs the e-commerce network community discovery results to e-commerce platform operators or researchers in related fields for product recommendations, store recommendations, friend or group chat recommendations, user interest analysis, and interest community analysis related tasks; the specific data processing and calculation work in all units is completed by the computer processor, and all units interact with the computer memory data.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] 1. The present invention provides an e-commerce network community discovery method based on semi-supervised network representation learning. By constructing self-supervisory information, it improves the guidance and supervision role in the user representation learning process, thereby enhancing the similarity between user nodes of the same type and the difference between user nodes of different types, which is conducive to providing richer discriminative information for e-commerce network community discovery.

[0054] 2. The present invention proposes an e-commerce network community discovery method based on semi-supervised network representation learning. By constructing a dual-channel network representation learning model, it effectively integrates self-supervised information constructed from e-commerce network user data and some existing user label information, thereby improving the ability to extract user features, enhancing the cohesion of user representations, and making the e-commerce network community discovery results more effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1This is a structural diagram of the e-commerce network community discovery model based on semi-supervised network representation learning in step S20 of the present invention;

[0056] Figure 2 This is a system structure diagram of an e-commerce network community discovery system based on semi-supervised network representation learning according to the present invention;

[0057] Figure 3 This is a flowchart of an e-commerce network community discovery method based on semi-supervised network representation learning described in the present invention. DETAILED DESCRIPTION

[0058] In order to further illustrate the technical solution of the present invention, the present invention is further described below with reference to the accompanying drawings and embodiments.

[0059] The e-commerce network community discovery method based on semi-supervised network representation learning described in the present invention is implemented by a computer program. Figure 3 The process shown here details a specific implementation of the technical solution proposed in this invention. A user shopping dataset was obtained from the JD.com shopping platform. This dataset contained 2,708 JD.com users, 270 of whom were categorized and had 5,429 friendship relationships between them. A 1,433-dimensional feature set was derived based on JD.com users' purchase data, browsing data, store follow data, event participation data, and product collection data. The technical solution of this invention utilizes user datasets captured from JD.com webpages for processing, enabling community discovery within the e-commerce network.

[0060] The implementation method mainly includes the following key contents:

[0061] S10: Obtain an e-commerce network user dataset, construct an e-commerce network structure based on the friendship relationships between users, use users' purchase data, browsing data, store attention data, activity participation data, and product collection data as user node features, use users of labeled types as label data, and construct an adjacency matrix, feature matrix, and label matrix based on the e-commerce network structure, user node features, and partial label data. The specific steps are as follows:

[0062] S11. Obtain and load the JD.com user dataset from the JD.com platform, build an e-commerce network structure based on the friendship between users, use the user's purchase data, browsing data, store attention data, and product collection data as user node features, and use the marked type of users as label data; denote the e-commerce network as G = (V, E, X), V = {v1, v2, ..., v N} represents the set of N = 2708 user nodes in the e-commerce network, where the nth user (1 n N) is denoted as user node v n ; It represents a set of L=270 user nodes with marked types in the e-commerce network, where the lth user (1 l L) is recorded as a user node with marked type E={e1,e2,…,e M} indicates that there are M = 5429 edges between user nodes, where the mth edge (1 m M) is denoted as e m ; X is an N×R dimensional user node feature matrix, whose nth (1 n N) row x n =[x n1 ,x n2 ,… x nR ] represents the user node v in the e-commerce network n The R=1433 features, where the element x nr Represents user node v n The rth feature of ; A represents an N×N adjacency matrix, corresponding to the M edges contained in the set E, and the element in the i-th row and j-th column (1 i N, 1 j N) in the matrix is ​​A ij Represents the i-th user node v i and the jth user node v j Adjacency relationship, if user node v i and user node v j If there is a connecting edge between them, then A ij =1, otherwise A ij =0; the user node category labeling in the e-commerce network is recorded through the L×K dimensional label matrix Y, where L represents the number of user nodes with category labels, K=7 represents the number of label categories, and the element in the i-th row and j-th column (1 i L, 1 j K) in the matrix is ​​Y ij Represents the user node of the i-th marked type The relationship between the j-th class label, if the user node of the marked type If it belongs to this category, then Y ij =1, otherwise Y ij =0.

[0063] S20. Based on the adjacency matrix, feature matrix and label matrix established in step S10, an e-commerce network community discovery model for semi-supervised network representation learning is constructed, which includes three parts: a semi-supervised channel, a self-supervised channel and a redundancy elimination module. Among them: the semi-supervised channel includes a GCN encoder, which uses the user information of the labeled type as a guide to learn the common features between users with the same interests and the difference features between users with different interests; the self-supervised channel also includes a GCN encoder, which constructs false user feature information and compares it with the user's real feature information, so that the user's low-dimensional representation can better capture and maintain the structure and characteristics of the e-commerce network under the real-false feature information comparison mechanism; the redundancy elimination module is used to remove the overlapping information in the above two channels, enhance the effectiveness of the guidance and supervision role in the user's low-dimensional representation learning process, so that the e-commerce network community discovery model obtains stable performance. Its overall structure is shown in the attached figure. Figure 1 As shown in the figure, the specific operation steps of the model are:

[0064] S21. In the semi-supervised channel, the e-commerce network G = (V, E, X) is input into the GCN encoder for encoding to obtain the representation vector of the user node in the e-commerce network. The formal representation of the encoding process is shown in formula (1):

[0065]

[0066] in, D is the degree matrix of the user node, user node v i The degree is expressed as the number of edges associated with the user node; W0 and W1 are learnable parameter matrices; H = {h1,h2…,h N} is the set of representation vectors of user nodes in the e-commerce network, h i Represents user node v i The low-dimensional representation of the user is obtained after encoding; the input here consists of two parts: the e-commerce network user node feature matrix X and the e-commerce network adjacency matrix A; the activation function Relu(·) in the encoder is defined by formula (2):

[0067]

[0068] When constructing the semi-supervised channel loss, the present invention uses a neural network classifier based on Softmax and cross entropy loss functions as a mutual information estimator to evaluate the correlation between the user node feature X and the partial user node label Y, so that the user low-dimensional representation H = {h1,h2…,h N} can learn more relevant information about the user node label Y; the semi-supervised channel optimization objective is shown in formula (3):

[0069]

[0070] Among them, x l Represents the features of the lth user node with label, y l represents the label value of the lth user node with a label, y′ represents the user label category, P(y′) is the prior distribution probability of the user label, f(x l ) represents the low-dimensional representation of the lth labeled user node obtained by the GCN encoder;

[0071] S22. In the self-supervisory channel, the e-commerce network user node feature matrix X and the e-commerce network adjacency matrix A are also used as input, and the user low-dimensional representation is obtained through the GCN encoder. Represents user node v i The low-dimensional representation of the user is obtained after encoding; when constructing the self-supervised channel loss, we use the Jensen-Shannon estimator to measure the low-dimensional representation of the user With user node v n The correlation between the user nodes and their adjacent nodes is measured from two aspects: the user's node feature X and the adjacency matrix A. The overall optimization goal of this channel is shown in formula (4):

[0072]

[0073] Among them, G i Indicates that the user node v i and its adjacent user nodes constitute the subgraph, Indicates the consistency of reconstruction of user node features, I(a ij ;w ij ) indicates the consistency of user structure reconstruction, Low-dimensional representation of users in e-commerce networks and The correlation, a ij Represents user node v i and v j The weight of the edge between them;

[0074] The consistency of user node feature reconstruction is calculated by formula (5):

[0075]

[0076] Among them, x j Represents user node v n And the feature information of the adjacent user nodes, x′ j It is a false user node feature generated by disrupting the adjacency matrix A of the user node, while keeping the feature matrix X unchanged, to generate the corresponding false user node feature information; is a discriminator used to discriminate between real user node features and generated false user node features, as shown in formula (6):

[0077]

[0078] Among them, Θ1 is a trainable score matrix, Where T represents the transposition operation and σ is the sigmoid function, as shown in formula (7):

[0079]

[0080] The consistency of the reconstruction of the user structure of the e-commerce network is calculated by (8):

[0081] I(a ij ;w ij )=a ij logw ij +(1-a ij )log(1-w ij ) (8)

[0082] S23, using the redundancy elimination module to identify and measure redundant information, the user low-dimensional representation obtained in the semi-supervised channel and the self-supervised channel is h l and The correlation between the two is used to measure information redundancy, as shown in formula (9):

[0083]

[0084] in, It is aimed at Generated false data, is the discriminator, h l T Where T is the transposition operation, as shown in formula (10):

[0085]

[0086] S30: For the e-commerce network community discovery model based on semi-supervised network representation learning constructed in step S20, a joint optimization objective function is designed, model training is performed by minimizing the objective function, and model parameters are determined, including the following specific steps:

[0087] S31, the semi-supervised channel, self-supervised channel and redundancy elimination module described in S20 are jointly trained to define the overall optimization objective function for the e-commerce network community discovery model based on semi-supervised network representation learning as shown in formula (11):

[0088]

[0089] in, is the semi-supervised channel loss, is the self-supervised channel loss, I r is redundant information, θ is a hyperparameter used to measure the importance of the self-supervisory channel;

[0090] S32. Initialize the parameters of the e-commerce network community discovery model using the semi-supervised network representation learning in step S20. Given an adjustment coefficient θ and a number of iterations T, iteratively execute steps S33 to S35 until the number of iterations is reached, thereby completing the training of the e-commerce network community discovery model using the semi-supervised network representation learning and obtaining the optimal parameters of the model.

[0091] S33, the e-commerce network user adjacency matrix A and feature matrix X obtained in S11 are input into the semi-supervised channel and the self-supervised channel respectively, and the semi-supervised e-commerce network representation H = {h1,h2…,h N} and self-supervised e-commerce network representation

[0092] S34, according to formula (3), formula (4) and formula (9) respectively get the semi-supervised channel target loss Self-supervised channel target loss and redundant information I r ;

[0093] S35. Input the loss obtained in S34 into formula (11) to obtain the total loss of the e-commerce network community discovery model of the semi-supervised network representation learning, perform backpropagation, and cyclically update the semi-supervised channel neural network parameters W0 and W1 and the self-supervised channel neural network parameters And the score matrices Θ1, Θ2.

[0094] S40, using the e-commerce network community discovery model constructed by the semi-supervised network representation learning in step S20 and the model parameters determined in step S30, to achieve community discovery and detection on the e-commerce network, and output the community discovery results, including the following specific steps:

[0095] S41, obtain the optimal parameters of the model through S33 to S35, input new e-commerce network user data, and obtain the community label of each user node through formula (12);

[0096]

[0097] Among them, c (i) represents the community label of the i-th user node, z (i) represents the representation of the i-th user node, z (i) Calculated by formula (13),

[0098]

[0099] μ (j) represents the centroid of the jth cluster, μ (j) Calculated by formula (14),

[0100]

[0101] S42. Output the e-commerce network community discovery results to e-commerce platform operators or researchers in related fields for tasks such as product recommendations, store recommendations, friend or group chat recommendations, user interest analysis, and interest community analysis.

[0102] Technical effect evaluation:

[0103] In order to verify the effectiveness and advancement of the technical solution proposed in this invention, the present invention is compared with several classic methods, including prediction-based methods (Planetoid), graph neural network-based methods (GCN, GWNN), and self-supervised learning methods (GMI). The average accuracy of 20 experiments, NMI, and ARI are used as evaluation indicators to compare and analyze the above matching results. The comparison results are shown in Table 1:

[0104] Table 1 Comparative analysis of results

[0105]

[0106] It can be seen from the results in the table that the technical solution of the present invention can achieve better accuracy, NMI and ARI values ​​when performing community discovery on e-commerce networks.

[0107] like Figure 2As shown, an e-commerce network community discovery system based on semi-supervised network representation learning includes a computer processor and memory, an e-commerce network adjacency matrix, a feature matrix and a label matrix unit, an e-commerce network community discovery model training unit based on semi-supervised network representation learning, and an e-commerce network community discovery model result output unit based on semi-supervised network representation learning; the e-commerce network adjacency matrix, feature matrix and label matrix unit execute step S10, for the acquired user data set, construct an e-commerce network structure according to the friendship relationship between users, use the user's purchase data, browsing data, store attention data, activity participation data and product collection data as user node features, and use the marked type of users as label data, thereby establishing the adjacency matrix, feature matrix and label matrix, and loading them into the computer memory; based on semi-supervised network representation learning, an e-commerce network community discovery system is shown. The e-commerce network community discovery model training unit based on network representation learning executes steps S20 to S30 according to the adjacency matrix, feature matrix and label matrix constructed in step S10, constructs an e-commerce network community discovery model based on semi-supervised network representation learning, and trains the model based on the defined objective function; the e-commerce network community discovery model result output unit based on semi-supervised network representation learning executes step S40, divides users with similar interests in the e-commerce network into the same community, and outputs the e-commerce network community discovery results to e-commerce platform operators or researchers in related fields for product recommendations, store recommendations, friend or group chat recommendations, user interest analysis, and interest community analysis related tasks; the specific data processing and calculation work in all units is completed by the computer processor, and all units interact with the computer memory data.

Claims

1. A method for discovering e-commerce network communities based on semi-supervised network representation learning, characterized by: The following steps are involved: S10. Obtain an e-commerce network user dataset, construct an e-commerce network structure based on the friendship relationships between users, use users' purchase data, browsing data, store attention data, event participation data, and product collection data as user node features, use users with labeled types as label data, and construct an adjacency matrix A, a feature matrix X, and a label matrix Y based on the e-commerce network structure, user node features, and partial label data; S20, based on the adjacency matrix, feature matrix, and label matrix established in step S10, using two independent GCN encoders to build an e-commerce network community discovery model with semi-supervised network representation learning; S30. For the e-commerce network community discovery model based on the semi-supervised network representation learning constructed in step S20, a joint optimization objective function is designed, and the model parameters are determined by training the model by minimizing the objective function. S40, using the e-commerce network community discovery model constructed by the semi-supervised network representation learning in step S20 and the model parameters determined in step S30 to achieve community discovery and detection on the e-commerce network, and output the community discovery results; The e-commerce network community discovery model of the semi-supervised network representation learning in step S20 includes three parts: a semi-supervised channel, a self-supervised channel, and a redundancy elimination module, wherein: the semi-supervised channel includes a GCN encoder, which uses user information of labeled types as a guide to learn the common features between users with the same interests and the differential features between users with different interests; the self-supervised channel also includes a GCN encoder, which constructs false user feature information and compares it with the user's real feature information, so that the user's low-dimensional representation can better capture and maintain the structure and characteristics of the e-commerce network under the real-false feature information comparison mechanism; the redundancy elimination module is used to remove the overlapping information in the above two channels, enhance the effectiveness of the guidance and supervision role in the user low-dimensional representation learning process, and thus enable the e-commerce network community discovery model to achieve stable performance; the specific operation steps of the model are: S21. In the semi-supervised channel, the e-commerce network G = (V, E, X) is input into the GCN encoder for encoding to obtain the representation vector of the user node in the e-commerce network. The formal representation of the encoding process is shown in formula (1): in, D is the degree matrix of the user node, user node v i The degree is expressed as the number of edges associated with the user node; W0 and W1 are learnable parameter matrices; H = {h1,h2…,h N } is the set of representation vectors of user nodes in the e-commerce network, h i Represents user node v i The low-dimensional representation of the user is obtained after encoding; the input here consists of two parts: the e-commerce network user node feature matrix X and the e-commerce network adjacency matrix A; the activation function Relu(·) in the encoder is defined by formula (2): When constructing the semi-supervised channel loss, a neural network classifier based on Softmax and cross entropy loss function is used as the mutual information estimator to evaluate the correlation between the user node feature X and the partial user node label Y, so that the user low-dimensional representation H = {h1,h2…,h N }Learn more relevant information about the user node label Y; the semi-supervised channel optimization objective is shown in formula (3): Among them, x l Represents the features of the lth user node with label, y l represents the label value of the lth user node with a label, y′ represents the user label category, P(y′) is the prior distribution probability of the user label, f(x l ) represents the low-dimensional representation of the lth labeled user node obtained by the GCN encoder; S22. In the self-supervisory channel, the e-commerce network user node feature matrix X and the e-commerce network adjacency matrix A are also used as input, and the user low-dimensional representation is obtained through the GCN encoder. Represents user node v i The low-dimensional representation of the user is obtained after encoding; when constructing the self-supervised channel loss, the Jensen-Shannon estimator is used to measure the low-dimensional representation of the user With user node v n The correlation between the user nodes and their adjacent nodes is measured from two aspects: the user's node feature X and the adjacency matrix A. The overall optimization goal of this channel is shown in formula (4): Among them, G i Indicates that the user node v i and its adjacent user nodes constitute the subgraph, Indicates the consistency of reconstruction of user node features, I(a ij ;w ij ) indicates the consistency of user structure reconstruction, Low-dimensional representation of users in e-commerce networks and The correlation, a ij Represents user node v i and v j The weight of the edge between them; The consistency of user node feature reconstruction is calculated by formula (5): Among them, x j Represents user node v n And the feature information of the adjacent user nodes, x′ j It is a false user node feature generated by disrupting the adjacency matrix A of the user node, while keeping the feature matrix X unchanged, to generate the corresponding false user node feature information; is a discriminator used to discriminate between real user node features and generated false user node features, as shown in formula (6): Among them, Θ1 is a trainable score matrix, Where T represents the transposition operation and σ is the sigmoid function, as shown in formula (7): The consistency of the reconstruction of the user structure of the e-commerce network is calculated by (8): I(a ij ;w ij )=a ij logw ij +(1-a ij )log(1-w ij ) (8) S23, using the redundancy elimination module to identify and measure redundant information, the user low-dimensional representation obtained in the semi-supervised channel and the self-supervised channel is h l and The correlation between the two is used to measure information redundancy, as shown in formula (9): in, It is aimed at Generated false data, is the discriminator, h l T Where T is the transposition operation, as shown in formula (10): The step S30 includes the following specific steps: S31, the semi-supervised channel, self-supervised channel and redundancy elimination module described in S20 are jointly trained to define the overall optimization objective function for the e-commerce network community discovery model based on semi-supervised network representation learning as shown in formula (11): in, is the semi-supervised channel loss, is the self-supervised channel loss, I r is redundant information, θ is a hyperparameter used to measure the importance of the self-supervisory channel; S32. Initialize the parameters of the e-commerce network community discovery model using the semi-supervised network representation learning in step S20. Given an adjustment coefficient θ and a number of iterations T, iteratively execute steps S33 to S35 until the number of iterations is reached, thereby completing the training of the e-commerce network community discovery model using the semi-supervised network representation learning and obtaining the optimal parameters of the model. S33, take the adjacency matrix A and feature matrix X constructed in S10 as input, input them into the semi-supervised channel and self-supervised channel respectively, and obtain the semi-supervised e-commerce network representation H = {h1,h2…,h N } and self-supervised e-commerce network representation S34, according to formula (3), formula (4) and formula (9) respectively get the semi-supervised channel target loss Self-supervised channel target loss and redundant information I r ; S35. Input the loss obtained in S34 into formula (11) to obtain the total loss of the e-commerce network community discovery model of the semi-supervised network representation learning, perform backpropagation, and cyclically update the semi-supervised channel neural network parameters W0 and W1 and the self-supervised channel neural network parameters And the score matrices Θ1, Θ2.

2. The e-commerce network community discovery method based on semi-supervised network representation learning according to claim 1 is characterized in that: The step S10 of the method comprises the following specific steps: S11. Obtain and load the e-commerce network user dataset from the e-commerce network platform, build the e-commerce network structure based on the friendship relationship between users, use the user's purchase data, browsing data, store attention data and product collection data as user node features, and use the marked type of users as label data; denote the e-commerce network as G = (V, E, X), V = {v1, v2, ..., v N } represents the set of N user nodes in the e-commerce network, where the nth user (1n N) is recorded as user node v n ; Represents a set of L user nodes of marked types in the e-commerce network, where the lth user (1lL) is recorded as a user node of marked type E={e1,e2,…,e M } represents the M edges between user nodes, where the mth edge (1m M) is denoted as e m ; X is an N′R-dimensional user node feature matrix, whose nth (1n N) row x n =[x n1, x n2, … x nR ] represents the user node v in the e-commerce network n R features of which element x nr Represents user node v n The rth feature of ; A represents an N′N adjacency matrix, corresponding to the M edges contained in the set E, and the element in the i-th row and j-th column (1i N, 1jN) in the matrix is ​​A ij Represents the i-th user node v i and the jth user node v j Adjacency relationship, if user node v i and user node v j If there is a connecting edge between them, then A ij =1, otherwise A ij =0; the user node category labeling situation in the e-commerce network is recorded through the L′K-dimensional label matrix Y, where L represents the number of user nodes with category labels, K represents the number of label categories, and the element in the i-th row and j-th column (1i L, 1j K) in the matrix is ​​Y ij Represents the user node of the i-th marked type The relationship between the j-th class label, if the user node of the marked type If it belongs to this category, then Y ij =1, otherwise Y ij =0.

3. The e-commerce network community discovery method based on semi-supervised network representation learning according to claim 1 is characterized in that: The step S40 of the method includes the following specific steps: S41, obtain the optimal parameters of the model through S33 to S35, input new e-commerce network user data, and obtain the community label of each user node through formula (12); Among them, c (i) represents the community label of the i-th user node, z (i) represents the representation of the i-th user node, z (i) Calculated by formula (13), μ (j) represents the centroid of the jth cluster, μ (j) Calculated by formula (14), S42. Output the e-commerce network community discovery results to e-commerce platform operators or researchers in related fields for tasks such as product recommendations, store recommendations, friend or group chat recommendations, user interest analysis, and interest community analysis.

4. An e-commerce network community discovery system based on semi-supervised network representation learning, characterized by: A method for discovering e-commerce network communities based on semi-supervised network representation learning, for implementing any one of claims 1-3, comprising a computer processor and memory, an e-commerce network adjacency matrix, a feature matrix, and a label matrix unit, an e-commerce network community discovery model training unit based on semi-supervised network representation learning, and an e-commerce network community discovery model result output unit based on semi-supervised network representation learning; the e-commerce network adjacency matrix, feature matrix, and label matrix unit executing step S10, for an acquired user data set, constructing an e-commerce network structure based on the friendship relationships between users, using the user's purchase data, browsing data, store attention data, activity participation data, and product collection data as user node features, and using the marked type of users as label data, thereby establishing an adjacency matrix, a feature matrix, and a label matrix, and loading them into the computer memory; The e-commerce network community discovery model training unit based on semi-supervised network representation learning executes steps S20 to S30 according to the adjacency matrix, feature matrix and label matrix constructed in step S10, constructs an e-commerce network community discovery model based on semi-supervised network representation learning, and trains the model based on the defined objective function; the e-commerce network community discovery model result output unit based on semi-supervised network representation learning executes step S40, divides users with similar interests in the e-commerce network into the same community, and outputs the e-commerce network community discovery results to e-commerce platform operators or researchers in related fields for product recommendations, store recommendations, friend or group chat recommendations, user interest analysis, and interest community analysis related tasks; the specific data processing and calculation work in all units is completed by the computer processor, and all units interact with the computer memory data.

Images