Node classification method and user classification method for billion-scale attribute network

Abstract

The application discloses a node classification method for a billion-level attribute network, which comprises the following steps: acquiring existing large-scale attribute network data and corresponding data information, and modeling and constructing a closed neighborhood adjacency matrix and an attribute matrix; constructing a sparse random matrix and projecting the attribute matrix to obtain an initial embedding matrix; multiplying the initial embedding matrix and the closed neighborhood adjacency matrix iteratively to obtain a network embedding matrix; quantizing the network embedding matrix by 1 bit to obtain a quantized network embedding matrix; performing a logical NOT operation and a splicing operation on the quantized network embedding matrix to obtain an extended embedding matrix; taking the extended embedding matrix as a training data set to train a model and obtain a node classification model; and using the node classification model to classify nodes of a large-scale attribute network to be classified. The application further discloses a user classification method comprising the node classification method for the billion-level attribute network. The application can complete node classification of a billion-level attribute network, has low resource occupation, and is high in classification efficiency.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a node classification method and a user classification method for attribute networks with a scale of hundreds of millions. Background Technology

[0002] Network node classification is an important component of computer science. It aims to categorize nodes in a network into different classes or labels to better understand the network's structure and function. Network node classification is significant for functions such as user classification and recommendation in social networks, as well as for research in bioinformatics, social sciences, and other fields.

[0003] Traditional node classification methods typically rely on analyzing the network topology, such as graph theory-based approaches. In recent years, with the development of deep learning technology, Graph Neural Networks (GNNs) have become a popular solution for handling network node classification problems. GNN-based network node classification schemes can effectively represent nodes by learning patterns in the network, thus achieving a complete classification of network nodes; moreover, these schemes have achieved significant results in network node classification tasks.

[0004] However, current network node classification schemes, including traditional schemes and GNN-based schemes, are designed for relatively small-scale networks. In the real world, however, networks such as social networks or e-commerce platforms are already on the scale of hundreds of millions of nodes; this scale presents significant computational and storage challenges for classifying network nodes.

[0005] To address the node classification problem in attribute networks with hundreds of millions of nodes, researchers have proposed several node classification algorithms for such networks, such as xGCN, SketchNE, LightNE2.0, and NPR. However, these solutions often require massive amounts of computational and memory resources, which severely limits their practical application.

[0006] With the rapid development of social networks, the number of users has surged, and the number of network nodes has long reached hundreds of millions. Faced with such a massive number of user nodes, existing user classification schemes for social networks either rely on massive computing and storage resources to classify users and then recommend content and services, or they use traditional, simpler network node classification schemes. This results in lower reliability of user classification results, thus affecting the quality of subsequent content and service recommendations. Summary of the Invention

[0007] One of the objectives of this invention is to provide a node classification method for attribute networks with hundreds of millions of nodes that consumes fewer resources and has higher classification efficiency.

[0008] The second objective of this invention is to provide a user classification method that includes the node classification method for attribute networks with hundreds of millions of nodes.

[0009] The node classification method for attribute networks with hundreds of millions of nodes provided by this invention includes the following steps:

[0010] S1. Obtain existing large-scale attribute network data and corresponding data information;

[0011] S2. Based on the data information obtained in step S1, model the large-scale attribute network and construct the corresponding closed neighborhood adjacency matrix and attribute matrix;

[0012] S3. Construct a sparse random matrix based on the data obtained in step S1;

[0013] S4. Based on the sparse random matrix obtained in step S3, the attribute matrix is ​​projected and dimensionality reduced to obtain the initial embedding matrix;

[0014] S5. Iteratively multiply the initial embedding matrix obtained in step S4 with the closed neighborhood adjacency matrix to obtain the network embedding matrix;

[0015] S6. Perform 1-bit quantization on the network embedding matrix obtained in step S5 to obtain the quantized network embedding matrix;

[0016] S7. Perform logical NOT and concatenation operations on the quantized network embedding matrix obtained in step S6 to obtain the extended embedding matrix;

[0017] S8. Use the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset, train the selected model, and obtain the node classification model.

[0018] S9. Using the node classification model obtained in step S8, classify the large-scale attribute network to be classified, and complete the node classification of the large-scale attribute network to be classified.

[0019] Step S2, which involves modeling the large-scale attribute network based on the data information obtained in step S1 and constructing the corresponding closed-neighbor adjacency matrix and attribute matrix, specifically includes the following steps:

[0020] Based on the data obtained in step S1, a large-scale attribute network is modeled.

[0021] Construct the adjacency matrix G' and attribute matrix A for each large-scale attribute network;

[0022] The adjacency matrix G of the closed neighborhood is constructed as G = G' + I, where I is the identity matrix.

[0023] Step S3, which involves constructing a sparse random matrix based on the data obtained in step S1, specifically includes the following steps:

[0024] The attributes corresponding to all nodes in each large-scale attribute network obtained in step S1 are combined to form an attribute set;

[0025] Based on the attribute set, a sparse random matrix R is constructed using the following formula:

[0026]

[0027] In the formula R ij Let be the element in the i-th row and j-th column of the sparse random matrix R; s is an intermediate variable, and n_f is the total number of elements in the attribute set;

[0028] The number of rows in the constructed sparse random matrix R is equal to the number of columns in the attribute matrix A, and the number of columns in the sparse random matrix R is equal to the target dimension K after dimensionality reduction.

[0029] Step S4, based on the sparse random matrix obtained in step S3, involves projecting and reducing the dimensionality of the attribute matrix to obtain the initial embedding matrix. This specifically includes the following steps:

[0030] The initial embedding matrix is ​​obtained by using the following formula, thus completing the dimensionality reduction of the attribute matrix through projection:

[0031] X (0) =A·R

[0032] In the formula X (0) This is the initial embedding matrix.

[0033] Step S5, which involves iteratively multiplying the initial embedding matrix obtained in step S4 with the adjacency matrix of the closed neighborhood to obtain the network embedding matrix, specifically includes the following steps:

[0034] The network embedding matrix is ​​obtained by iteratively multiplying the initial embedding matrix and the closed neighborhood adjacency matrix using the following formula:

[0035] X (TT) =G TT ·A·R

[0036] In the formula X (TT) is the network embedding matrix; TT is the number of iterations of the closed neighborhood adjacency matrix multiplication.

[0037] Step S6, which involves performing 1-bit quantization on the network embedding matrix obtained in step S5 to obtain the quantized network embedding matrix, specifically includes the following steps:

[0038] For network embedding matrix X (TT) The element in the i-th row and j-th column The following rules are used for 1-bit quantization:

[0039] like If the value is greater than 0, then The value is directly corrected to 1;

[0040] like If the value is less than or equal to 0, then The value is directly corrected to 0;

[0041] For the network embedding matrix X (TT) After all elements are quantized to 1 bit, the quantized network embedding matrix H is obtained.

[0042] Step S7 involves performing a logical NOT operation and a concatenation operation on the quantized network embedding matrix obtained in step S6 to obtain the extended embedding matrix. Specifically, this includes the following steps:

[0043] The aforementioned logical NOT operation and concatenation operation specifically include the following steps: For the i-th row of the quantized network embedding matrix H, perform an inversion operation on each element in the row and record the inversion result; place the inversion result directly after the last element of the row to complete the concatenation operation;

[0044] For each row of the quantized network embedding matrix H, a logical NOT operation and a concatenation operation are performed to obtain the extended embedding matrix.

[0045] Step S8, which involves using the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset to train the selected model and obtain the node classification model, specifically includes the following steps:

[0046] Use the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset;

[0047] Select a model for node classification;

[0048] The selected model is trained using the training dataset; during training, a defined loss function is used.

[0049] During training, the parameters of the selected model are updated using the set optimization algorithm;

[0050] Finally, the trained model is obtained, which serves as the final node classification model.

[0051] Step S9 describes using the node classification model obtained in step S8 to classify the large-scale attribute network to be classified, specifically including the following steps:

[0052] For large-scale attribute networks to be classified, obtain the corresponding data information;

[0053] Based on the acquired data, the extended embedding matrix of the large-scale attribute network to be classified is calculated using the technical solutions in steps S2 to S7.

[0054] The extended embedding matrix of the large-scale attribute network to be classified is input into the node classification model obtained in step S8 and the corresponding classification result is obtained, thus completing the node classification of the large-scale attribute network to be classified.

[0055] This invention also provides a user classification method that includes the node classification method for attribute networks with hundreds of millions of nodes, specifically including the following steps:

[0056] A. Obtain data information about the social networks to be classified;

[0057] B. Based on the data information obtained in step A, the node classification method for attribute networks with hundreds of millions of nodes is used to classify the nodes of the social network to be classified.

[0058] C. Based on the node classification results obtained in step B, complete the user classification of the social network.

[0059] The node and user classification methods provided by this invention for attribute networks with hundreds of millions of nodes utilize a network embedding matrix calculated from randomly generated sparse random matrices and constructed closed-neighbor adjacency matrices to represent such networks. Simultaneously, a training dataset is formed through logical NOT and concatenation operations to train the classification model. Therefore, this invention avoids complex mathematical calculations and massive parameter learning, achieving node classification of attribute networks with hundreds of millions of nodes while maintaining accuracy, and also consuming fewer resources and achieving higher node classification efficiency. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the node classification method of the present invention.

[0061] Figure 2 This is a schematic diagram of the user classification method of the present invention. Detailed Implementation

[0062] like Figure 1 The diagram shows a flowchart of the node classification method of the present invention: This node classification method for attribute networks with hundreds of millions of nodes disclosed in the present invention includes the following steps:

[0063] S1. Obtain existing large-scale attribute network data and corresponding data information;

[0064] S2. Based on the data obtained in step S1, model the large-scale attribute network and construct the corresponding closed-neighbor adjacency matrix and attribute matrix; specifically including the following steps:

[0065] Based on the data obtained in step S1, a large-scale attribute network is modeled.

[0066] Construct the adjacency matrix G' and attribute matrix A for each large-scale attribute network;

[0067] Construct the adjacency matrix G of the closed neighborhood as G = G' + I, where I is the identity matrix;

[0068] S3. Construct a sparse random matrix based on the data obtained in step S1; specifically, this includes the following steps:

[0069] The attributes corresponding to all nodes in each large-scale attribute network obtained in step S1 are combined to form an attribute set;

[0070] Based on the attribute set, a sparse random matrix R is constructed using the following formula:

[0071]

[0072] In the formula R ij Let be the element in the i-th row and j-th column of the sparse random matrix R; s is an intermediate variable, and n_f is the total number of elements in the attribute set;

[0073] The number of rows in the constructed sparse random matrix R is equal to the number of columns in the attribute matrix A, and the number of columns in the sparse random matrix R is equal to the target dimension K after dimensionality reduction.

[0074] In practice, when calculating the values ​​of elements in a sparse random matrix R... The number to be multiplied (by 1, 0, or -1) is random, requiring only that among all elements of the sparse random matrix R, the probability of multiplying by 1 and multiplying by -1 is... The probability of multiplying by 0 is That's all;

[0075] S4. Based on the sparse random matrix obtained in step S3, the attribute matrix is ​​projected and its dimensionality reduced to obtain the initial embedding matrix; specifically, the following steps are included:

[0076] The initial embedding matrix is ​​obtained by using the following formula, thus completing the dimensionality reduction of the attribute matrix through projection:

[0077] X (0) =A·R

[0078] In the formula X (0) This is the initial embedding matrix;

[0079] The projection process in this step is mainly to efficiently reduce the dimensionality of networks with a scale of hundreds of millions while ensuring the embedding quality, because the value of the target dimension K is much smaller than the dimension of the attribute matrix A.

[0080] S5. Iteratively multiply the initial embedding matrix obtained in step S4 with the adjacency matrix of the closed neighborhood to obtain the network embedding matrix; specifically including the following steps:

[0081] The network embedding matrix is ​​obtained by iteratively multiplying the initial embedding matrix and the closed neighborhood adjacency matrix using the following formula:

[0082] X (TT) =G TT ·A·R

[0083] In the formula X (TT) is the network embedding matrix; TT is the number of iterations of multiplying the closed neighborhood adjacency matrix;

[0084] The purpose of this iterative multiplication step is to preserve the information of the node itself when exploring higher-order nodes;

[0085] S6. Perform 1-bit quantization on the network embedding matrix obtained in step S5 to obtain the quantized network embedding matrix; specifically, this includes the following steps:

[0086] For network embedding matrix X (TT) The element in the i-th row and j-th column The following rules are used for 1-bit quantization:

[0087] like If the value is greater than 0, then The value is directly corrected to 1;

[0088] like If the value is less than or equal to 0, then The value is directly corrected to 0;

[0089] For the network embedding matrix X (TT) After all elements are quantized to 1 bit, the quantized network embedding matrix H is obtained;

[0090] This step of 1-bit quantization can save a lot of storage space;

[0091] S7. Perform logical NOT and concatenation operations on the quantized network embedding matrix obtained in step S6 to obtain the extended embedding matrix; specifically including the following steps:

[0092] The aforementioned logical NOT operation and concatenation operation specifically include the following steps: For the i-th row of the quantized network embedding matrix H, perform an inversion operation on each element in the row and record the inversion result; place the inversion result directly after the last element of the row to complete the concatenation operation;

[0093] For each row of the quantized network embedding matrix H, a logical NOT operation and a concatenation operation are performed to obtain the extended embedding matrix.

[0094] In practical implementation, the value of the quantized network embedding matrix H is taken as follows: Let's take an example to illustrate:

[0095] For the first row of the quantized network embedding matrix H, the element [1 1 0 1] is first inverted and recorded to obtain [0 0 1 0]; then [0 0 1 0] is directly placed after the last element of [1 1 0 1] to obtain [1 10 1 0 0 1 0].

[0096] Similarly, for the second row of the quantized network embedding matrix H, the elements [0 0 1 0] are first inverted and recorded to obtain [1 1 0 1]; then [1 1 0 1] is directly placed after the last element of [0 0 1 0] to obtain [0 0 1 0 1 1 0 1].

[0097] Finally, the extended embedding matrix is ​​obtained as follows:

[0098] This step bypasses large-scale kernel matrix storage, thereby further reducing resource consumption.

[0099] S8. Using the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset, train the selected model to obtain the node classification model; specifically including the following steps:

[0100] Use the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset;

[0101] Select a model for node classification;

[0102] The selected model is trained using the training dataset; during training, a defined loss function is used.

[0103] During training, the parameters of the selected model are updated using the set optimization algorithm;

[0104] Finally, the trained model is obtained as the final node classification model;

[0105] In practice, the models used for node classification can include logistic regression, support vector machine, K-nearest neighbor, and neural network models.

[0106] When using a logistic regression model, the specific details of this step are as follows:

[0107] Use the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset;

[0108] The model selected for node classification is logistic regression. Logistic regression is a generalized linear model used to predict the probability of binary or multi-class classification problems. For node i, the model predicts the probability that the node belongs to a certain class, as shown in the formula:

[0109]

[0110] In the formula, P = (Y = i|x) represents the probability that a node belongs to category i given feature x; W i and b i These are the weight vector and bias value corresponding to category i, respectively; M is the total number of categories;

[0111] The selected model is trained using the training dataset; during training, a defined loss function is used: the loss function is... Where N is the total number of samples; y ij y is an indicator variable; if node i belongs to category j, then y ij =1, otherwise 0; log is the natural logarithm;

[0112] During training, the parameters of the selected model are updated using a predefined optimization algorithm (such as gradient descent). The goal is to minimize the loss function and find the optimal weights W and biases b so that the model can accurately predict the category of the node.

[0113] S9. Using the node classification model obtained in step S8, classify the large-scale attribute network to be classified, and complete the node classification of the large-scale attribute network to be classified; specifically including the following steps:

[0114] For large-scale attribute networks to be classified, obtain the corresponding data information;

[0115] Based on the acquired data, the extended embedding matrix of the large-scale attribute network to be classified is calculated using the technical solutions in steps S2 to S7.

[0116] The extended embedding matrix of the large-scale attribute network to be classified is input into the node classification model obtained in step S8 and the corresponding classification result is obtained, thus completing the node classification of the large-scale attribute network to be classified.

[0117] This invention represents a billion-level attribute network by iteratively multiplying a sparse matrix with a large number of zero-valued terms with an adjacency matrix with a closed neighborhood to obtain an embedding matrix. This not only significantly improves the computation speed and reduces memory consumption efficiency of billion-level attribute networks, but also preserves the structure and node attribute information of the attribute network, reducing the dimensionality of the data while maintaining similarity. This invention eliminates the need to generate a kernel matrix of size squared for the number of training instances for support vector machines. Instead, it expands the matrix using a logical NOT operation to form an inner product space, and inputs the expanded embedding matrix into linear learning algorithms such as logistic regression for training and testing, thereby effectively classifying various types of nodes. Therefore, this invention avoids the complex mathematical calculations and massive parameter learning required in neural networks, and no longer relies on expensive high-end hardware. It can significantly reduce time and space overhead while meeting node classification accuracy, making it particularly suitable for node classification tasks in billion-level attribute network scenarios. Moreover, this invention has high reliability, good accuracy, high efficiency, and low resource consumption.

[0118] The node classification method of the present invention will be described below with reference to an embodiment:

[0119] This embodiment uses publicly available datasets for node classification testing; these datasets include the OGBN-Product dataset, the Amazon dataset, the MAG-Scholar-C dataset, and the OGBN-Papers100M dataset. The OGBN-Product dataset contains 2,449,029 nodes representing products across 47 classes, and 61,859,140 edges, indicating that these products were purchased together on the Amazon platform. Each product is represented as a 100-dimensional feature vector. Amazon collected 1,137,822 nodes from the same user as products, and 1,010,566,473 edges as reviews. Each product is represented as a 1000-dimensional feature vector, and a total of 99 labels are assigned to each product. MAG-Scholar-C consists of 10,541,560 papers across 8 subject areas from the Microsoft Academic Graph. Each paper is represented as a 2,784,240-dimensional feature vector, containing a total of 132,609,997 citation relationships. The ogbn-papers100M dataset consists of 111,059,956 nodes representing papers and 1,615,685,872 edges representing citation relationships in the Microsoft Scholar graph. Each paper is represented as a 128-dimensional feature vector. Only 1,546,782 papers on arXiv have labels spanning 172 subject areas. Detailed descriptions of the ogbn-product dataset, Amazon dataset, MAG-Scholar-C dataset, and ogbn-papers100M dataset are shown in Table 1.

[0120] Table 1. Dataset Description Table

[0121] Number of nodes 2,449,029 1,137,822 10,541,560 111,059,956 Number of sides 61,859,140 1,010,566,473 132,609,997 1,615,685,872 feature 100 1000 2,784,240 128 category 47 99 8 172

[0122] Then, the existing RandNE classification method, FastRP classification method, XGCN classification method, PANE++ classification method and the classification method of this application are used to classify the nodes in the above public datasets respectively;

[0123] Among them, RandNE approximates higher-order structural proximity through iterative Gaussian random projection of the traditional adjacency matrix. FastRP improves the efficiency of RandNE by replacing Gaussian random projection with sparse random projection. xGCN only retains the structural information in the node embedding, which is initialized according to message propagation and then optimized in RefNet. PANE++ incorporates attribute and structural information into the forward and reverse affinity random walks. The network embedding representations obtained from the above algorithms are input into a linear learner for training and testing, thereby completing the node classification experiment. After classification, the performance data of the corresponding classification methods are shown in Table 2:

[0124] Table 2. Comparison of Classification Method Performance Data

[0125]

[0126] As shown in Table 2, on these four different datasets, the proposed method significantly outperforms all common network embedding algorithms (i.e., RandNE, FastRP, and xGCN); furthermore, it competes well with attribute network embedding methods (i.e., PANE++), demonstrating the promising potential of quantized random projection in preserving patterns in billion-scale attribute networks without data distribution fitting. PANE++ performs much better than OGBN-product on MAG-Scholar-C because it is designed for networks with large attribute sets.

[0127] Table 2 also shows the time and space efficiency of the embedding process. The method in this application executes much faster than the learning methods (i.e., xGCN and PANE++), with a speedup of up to 66 times, and is also generally superior in terms of memory consumption. xGCN is highly dependent on GPU memory, so it fails on the two largest datasets (MAG-Scholar-C and ogbn-papers100M).

[0128] PANE++ ran out of time on Amazon and OGBN-Papers100M, two home networks with billions of edges. Although our method introduces an additional attribute matrix, it competes well with random projection methods (i.e., RandNE and FastRP) in terms of time and space cost, as the latter requires multiple projection matrices and generates 64-bit node representations. In particular, FastRP runs out of memory on OGBN-Papers100M when T>3, meaning it cannot capture higher-order nearest neighbors. Overall, our method achieves a good balance between accuracy and time and space cost.

[0129] like Figure 2 The diagram shown is a flowchart of the user classification method of the present invention: The user classification method disclosed in this invention, which includes the node classification method for attribute networks with hundreds of millions of nodes, specifically includes the following steps:

[0130] A. Obtain data information about the social networks to be classified; specifically including the following steps:

[0131] Collect user behavior data on the platform, including content posted, likes, comments, follows, shares, etc.; at the same time, collect users' personal information, including gender, age, geolocation, etc.

[0132] B. Based on the data obtained in step A, the node classification method for attribute networks with hundreds of millions of nodes is used to classify the nodes of the social network to be classified; specifically, the following steps are included:

[0133] Construct an adjacency matrix and an attribute matrix from user behavior and profile information;

[0134] Based on the constructed adjacency matrix and attribute matrix, the nodes of the social network are classified using the node classification method for attribute networks with hundreds of millions of nodes.

[0135] Based on the node classification results, users are divided into different groups; at this point, each group represents a set of users with similar behaviors and interests.

[0136] C. Based on the node classification results obtained in step B, complete the user classification of the social network. At this point, the user classification results indicate that social network users in the same category have the same or similar user characteristics and user interests. Therefore, based on the user classification results, subsequent content recommendations, recommendations of users with similar interests, etc., can be completed to improve the user experience and the quality of social interaction on the social network.

[0137] Furthermore, the node classification method for attribute networks with hundreds of millions of nodes provided by this invention is also applicable to product classification, user classification and product recommendation on e-commerce platforms, user classification and course recommendation on online education platforms, user classification, product classification and product recommendation in the smart home field, and human resource management, among other applications.

Claims

1. A user classification method, characterized in that... Specifically, the steps include the following: A. Obtain data information about the social networks to be classified; specifically including the following steps: Collect user behavior data on the platform, including content posted, likes, comments, follows, and shares; also collect user profile information, including gender, age, and geographic location; B. Based on the data obtained in step A, a node classification method for attribute networks with hundreds of millions of nodes is used to classify the nodes of the social network to be classified; specifically, the following steps are included: Construct an adjacency matrix and an attribute matrix from user behavior and profile information; Based on the constructed adjacency matrix and attribute matrix, the nodes of the social network are classified using the node classification method for attribute networks with hundreds of millions of nodes. Based on the node classification results, users are divided into different groups; at this point, each group represents a set of users with similar behaviors and interests. C. Based on the node classification results obtained in step B, complete the user classification of the social network; at this point, the user classification results indicate that social network users in the same category have the same or similar user characteristics and user interests. In practical implementation, the node classification method for attribute networks with hundreds of millions of nodes includes the following steps: S1. Obtain existing large-scale attribute network data and corresponding data information; S2. Based on the data information obtained in step S1, model the large-scale attribute network and construct the corresponding closed neighborhood adjacency matrix and attribute matrix; S3. Construct a sparse random matrix based on the data obtained in step S1; S4. Based on the sparse random matrix obtained in step S3, the attribute matrix is ​​projected and dimensionality reduced to obtain the initial embedding matrix; S5. Iteratively multiply the initial embedding matrix obtained in step S4 with the closed neighborhood adjacency matrix to obtain the network embedding matrix; S6. Perform 1-bit quantization on the network embedding matrix obtained in step S5 to obtain the quantized network embedding matrix; S7. Perform logical NOT and concatenation operations on the quantized network embedding matrix obtained in step S6 to obtain the extended embedding matrix; S8. Use the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset, train the selected model, and obtain the node classification model. S9. Using the node classification model obtained in step S8, classify the large-scale attribute network to be classified, and complete the node classification of the large-scale attribute network to be classified.

2. The user classification method according to claim 1, characterized in that... Step S2, which involves modeling the large-scale attribute network based on the data information obtained in step S1 and constructing the corresponding closed-neighbor adjacency matrix and attribute matrix, specifically includes the following steps: Based on the data obtained in step S1, a large-scale attribute network is modeled. Construct the adjacency matrix of each large-scale attribute network and attribute matrix ; Constructing a closed neighborhood adjacency matrix for ,in It is an identity matrix.

3. The user classification method according to claim 2, characterized in that... Step S3, which involves constructing a sparse random matrix based on the data obtained in step S1, specifically includes the following steps: The attributes corresponding to all nodes in each large-scale attribute network obtained in step S1 are combined to form an attribute set; Based on the attribute set, a sparse random matrix is ​​constructed using the following formula. : In the formula For sparse random matrices The element in the i-th row and j-th column; It is an intermediate variable, and , The total number of elements in the attribute set; Constructed sparse random matrix The number of rows is the attribute matrix The number of columns, sparse random matrix The number of columns is the target dimension after dimensionality reduction. .

4. The user classification method according to claim 3, characterized in that... Step S4, based on the sparse random matrix obtained in step S3, involves projecting and reducing the dimensionality of the attribute matrix to obtain the initial embedding matrix. This specifically includes the following steps: The initial embedding matrix is ​​obtained by using the following formula, thus completing the dimensionality reduction of the attribute matrix through projection: In the formula This is the initial embedding matrix.

5. The user classification method according to claim 4, characterized in that... Step S5, which involves iteratively multiplying the initial embedding matrix obtained in step S4 with the closed neighborhood adjacency matrix to obtain the network embedding matrix, specifically includes the following steps: The network embedding matrix is ​​obtained by iteratively multiplying the initial embedding matrix and the closed neighborhood adjacency matrix using the following formula: In the formula For network embedding matrix; denoted as the number of iterations of the adjacency matrix of the closed neighborhood.

6. The user classification method according to claim 5, characterized in that... Step S6, which involves performing 1-bit quantization on the network embedding matrix obtained in step S5 to obtain the quantized network embedding matrix, specifically includes the following steps: For network embedding matrix The element in the i-th row and j-th column The following rules are used for 1-bit quantization: like If the value is greater than 0, then The value is directly corrected to 1; like If the value is less than or equal to 0, then The value is directly corrected to 0; Network embedding matrix After all elements are quantized to 1 bit, the quantized network embedding matrix is ​​obtained. .

7. The user classification method according to claim 6, characterized in that... Step S7 involves performing a logical NOT operation and a concatenation operation on the quantized network embedding matrix obtained in step S6 to obtain the extended embedding matrix. Specifically, this includes the following steps: The aforementioned logical NOT and concatenation operations specifically include the following steps: For the quantized network embedding matrix... For the element in the i-th row, invert each element in the row and record the inversion result; place the inverted result directly after the last element in the row to complete the concatenation operation; Quantization network embedding matrix Each row is subjected to a logical NOT operation and a concatenation operation, resulting in an extended embedding matrix.

8. The user classification method according to claim 7, characterized in that... Step S8, which involves using the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset to train the selected model and obtain the node classification model, specifically includes the following steps: Use the extended embedding matrices corresponding to all large-scale attribute networks as the training dataset; Select a model for node classification; The selected model is trained using the training dataset; during training, a defined loss function is used. During training, the parameters of the selected model are updated using the set optimization algorithm; Finally, the trained model is obtained, which serves as the final node classification model.

9. The user classification method according to claim 8, characterized in that... Step S9 describes using the node classification model obtained in step S8 to classify the large-scale attribute network to be classified, specifically including the following steps: For large-scale attribute networks to be classified, obtain the corresponding data information; Based on the acquired data, the extended embedding matrix of the large-scale attribute network to be classified is calculated using the technical solutions in steps S2 to S7. The extended embedding matrix of the large-scale attribute network to be classified is input into the node classification model obtained in step S8 and the corresponding classification result is obtained, thus completing the node classification of the large-scale attribute network to be classified.

Images