A User Risk Classification Method Based on Particle Topology Construction and Feature Enhancement

Abstract

This invention discloses a user risk classification method based on granular topology construction and feature enhancement, comprising: after acquiring user multi-view behavior data, performing unsupervised granular sphere partitioning of user nodes under each view; constructing an initial topology based on intra- and inter-granular connections; and suppressing noisy edges through a differentiable shrinkage activation mechanism to obtain a purified user relationship topology; simultaneously, granular sphere partitioning and refinement of feature dimensions of each view; concatenating joint features and generating a weight matrix through a multilayer perceptron to enhance the original features element-wise; performing information propagation on the enhanced features based on the purified topology; fusing the embeddings of each view to obtain a unified node representation; and outputting the user risk category through a risk classifier. This invention effectively suppresses noisy relationships introduced by accidental similarity and spoofing behavior, fully utilizes complementary information from multiple views, significantly improves the accuracy and robustness of risk user identification, and is suitable for practical risk control scenarios with scarce labels.

Description

Technical Field

[0001] This invention relates to the field of user risk classification technology, and in particular to a user risk classification method based on particle topology construction and feature enhancement. Background Technology

[0002] In today's internet platforms, financial service platforms, e-commerce platforms, and content communities, user risk is a crucial element in ensuring the safe operation and business stability of these platforms. As risky behaviors such as abnormal registrations, fraudulent transactions, coupon abuse, malicious marketing, bulk order placement, and other violations continue to evolve, risky users often possess strong concealment and disguise capabilities. Graph learning methods, due to their ability to characterize the relationships and similarities between user accounts, have been increasingly applied to user risk identification tasks. Existing methods typically treat user accounts as graph nodes, using shared devices, shared IP addresses, transaction associations, or behavioral similarities among users as edges, and classifying user node risks through neighborhood information propagation.

[0003] However, in real-world business scenarios, user relationship topologies are often unreliable. Normal users and risky users may form accidental connections due to shared network environments, similar transaction times, the same delivery areas, or similar browsing paths. Risky users may also impersonate others by changing devices, IP addresses, transaction paths, or account associations. If information is propagated directly over such noisy topologies, the model is prone to aggregating interfering information from irrelevant or impersonating users, leading to the smoothing and weakening of risky user characteristics. Furthermore, user behavior data typically exhibits multi-view characteristics. Different views reflect user risk status from perspectives such as registration information, login devices, network environment, transaction behavior, browsing and clicking behavior, and account association information. However, if complementary information between different views and related features within a view are not fully utilized, it will also affect the effectiveness of user node risk classification.

[0004] Existing solutions mainly include reconstructing the user graph based on a fixed nearest neighbor strategy and using multi-view feature stitching for risk classification. However, these methods still have significant shortcomings: First, topology construction relies on fixed nearest neighbor parameters, making it difficult to adapt to the relationship distribution of different platforms, different business stages, and different user groups, and easily introducing noisy user relationship edges; second, the multi-view feature fusion method is relatively simple, making it difficult to fully utilize the complementary information between different user behavior views. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a monitoring method for a dual-host status monitoring system based on multiple sensors, thereby solving the technical problems of existing technologies where topology construction relies on fixed nearest neighbor parameters and the multi-view feature fusion method is relatively simple.

[0006] This invention provides a user risk classification method based on particle topology construction and feature enhancement, comprising the following steps:

[0007] Step 1: Obtain user behavior network data and construct it into multi-view user node data, where each user account is a user node and the network data of each user are feature matrices under different views;

[0008] Step 2: Perform unsupervised granular partitioning on user nodes under each view, and construct the initial user topology under the corresponding view based on the topology within and between granular spheres; suppress noisy user relationship edges through a differentiable shrinking activation mechanism to obtain a purified user relationship topology;

[0009] Step 3: Refine the features within each view by using feature-level granular sphere partitioning, and then concatenate the refined user feature representations of each view into a user joint feature;

[0010] Step 4: Generate a weight matrix using a multilayer perceptron based on the user joint features. Then, use the weight matrix to perform element-wise weighted enhancement on the original user feature representations of each view to obtain an enhanced user feature representation that integrates complementary information from the views.

[0011] Step 5: Based on the purified user relationship topology, perform information propagation on the enhanced user feature representation to obtain the user node embedding representation under each view, and fuse the user node embedding representation of each view according to the learnable view fusion weight to obtain a unified user node representation;

[0012] Step 6: Use a risk classifier to predict user risk by representing the unified user nodes, and obtain user risk classification.

[0013] Furthermore, in step 2, the method for obtaining the topology within the spheres during the unsupervised sphere partitioning process is as follows:

[0014] Set the N user nodes in the current view as the initial spheres, so that... As a stopping condition for particle splitting, k-means++ is used to set the sphere center and divide it into sub-spheres, obtaining a set of user node-level particles. A connection is established between any two user nodes within the same particle to generate the intra-sphere topology. The specific formula for the user node-level particle set is as follows:

[0015] ;

[0016] In the formula, GB represents the first... The spheres generated in the nth view; m is the nth view. The number of user node-level particles generated under each view;

[0017] The internal topology of the generated spheres is as follows:

[0018] ;

[0019] In the formula, and The i-th and j-th user nodes are respectively in the i-th and j-th positions. Feature representation under each view; For the first The I-th particle in each view.

[0020] Furthermore, in step 2, the method for obtaining the topology between particles during the unsupervised particle partitioning process is as follows:

[0021] Calculate the representative node for each sphere, determine similar spheres based on the Euclidean distance between the representative nodes of different spheres, and generate the sphere topology by establishing user node connections between similar spheres.

[0022] The representative user node is the mean of the feature representations of all user nodes within the current sphere, and the specific formula is as follows:

[0023] ;

[0024] In the formula, Let be the representative user node of the i-th sphere; This represents the number of user nodes in the current sphere.

[0025] The formula for calculating the Euclidean distance between nodes represented by different spheres is as follows:

[0026] ;

[0027] In the formula, The distance is Euclidean. Let be the representative user node of the i-th sphere; Let J be the representative user node of the J-th sphere;

[0028] The resulting intergranular topology is as follows:

[0029] ;

[0030] In the formula, and The i-th and j-th user nodes are respectively in the i-th and j-th positions. Feature representation under each view and The first The I-th and J-th particles under each view.

[0031] Furthermore, in step 2, the method for constructing the initial user topology is specifically as follows:

[0032] The intra-sphere topology and inter-sphere topology are merged to obtain the initial user topology, as shown in the following formula:

[0033] ;

[0034] In the formula, For the first Topology of the sphere under each view; For the first The topology between particles and spheres under each view.

[0035] Furthermore, in step 2, the specific method for suppressing noisy user relationship edges through a differentiable shrinkable activation mechanism to obtain a purified user relationship topology is as follows:

[0036] The importance of edge connections between user nodes in the initial user topology is learned through a differentiable shrinkable activation mechanism, and noisy user relationship edges introduced by accidental similarity, anomalous behavior masquerading, and / or weak correlation features are suppressed. The specific formula is as follows:

[0037] ;

[0038] In the formula, The learnable edge coefficient matrix; The learnable threshold matrix; This is element-wise multiplication; It is a linear function; This is the initial user topology.

[0039] Furthermore, in step 3, the specific method for feature-level particle division is as follows:

[0040] For each user view, the feature dimensions in that view are used as the partitioning object for unsupervised granular sphere partitioning, resulting in multiple feature-level granular spheres. The specific formula is as follows:

[0041] ;

[0042] In the formula, For the first The set of feature-level particles under each user view; m is the number of feature-level particles.

[0043] Furthermore, in step 3, the specific method for refining the features within the view is as follows:

[0044] Construct a feature similarity matrix; use the feature similarity matrix to refine the original user feature representation in the current view to obtain a refined user feature representation. The specific method for constructing the feature similarity matrix is ​​as follows:

[0045] To determine whether any two user feature dimensions belong to the same feature level sphere, if they do, calculate their cosine similarity; if they do not, set the corresponding similarity to zero. The specific formula for constructing the feature similarity matrix is ​​as follows:

[0046] ;

[0047] In the formula, For the first Feature similarity matrix of each user view; and The first The feature vectors corresponding to the i-th and j-th feature dimensions in each user view; Cosine similarity; For the first The i-th particle in a user view;

[0048] The specific formula for refining features within a view is as follows:

[0049] ;

[0050] In the formula, For the first Refined user feature representation under each user view;

[0051] The specific formula for user joint features is:

[0052] ;

[0053] middle, For feature splicing; For user joint feature representation; represents the feature dimensions of different views; N is the number of user nodes; R is the real number field.

[0054] Furthermore, in step 4, the specific formula for obtaining the weight matrix is ​​as follows:

[0055] ;

[0056] In the formula, For the first Weight matrix corresponding to each user view; For the first A multilayer perceptron corresponds to each user view; N is the number of user nodes; For the first Feature dimensions of each view;

[0057] The specific formula for obtaining the enhanced user feature representation is as follows:

[0058] ;

[0059] In the formula, For the first Enhanced feature representation under each user view; This is element-wise multiplication.

[0060] Furthermore, in step 5, the specific method for performing information dissemination is as follows:

[0061] The purified user relationship topology and enhanced user feature representation under each view are used as input to the graph convolutional network. Based on the purified user relationship topology, the neighborhood range of user nodes is determined, and the feature information of neighboring user nodes is aggregated to obtain the user node embedding representation under the corresponding view.

[0062] The specific formula for the user node embedding representation is as follows:

[0063] ;

[0064] In the formula, ;

[0065] For the first Node embedding under each user view, where... ; For activation functions; and The first The first view The learnable weight matrix of layer l and layer l; For the first The fusion weight of each user view.

[0066] Furthermore, in step 6, the prediction probability matrix of the risk classifier is:

[0067] ;

[0068] In the formula, H represents the embedding representation of the user node;

[0069] The classification loss is:

[0070] ;

[0071] In the formula, The set of labeled user nodes; Let i be the true label of the i-th user node belonging to the c-th class; This represents the probability that the i-th user node is predicted to be of class c.

[0072] The beneficial effects of this invention are:

[0073] This invention does not rely on the original user relationships that may contain noisy connections. Instead, it actively constructs user relationship topology based on the similarity of user behavior in multiple views and suppresses unreliable connections. This reduces interference connections between normal users and risky users caused by shared devices, similar network environments, the same delivery address, or accidental similar behavior, and more accurately identifies user groups with similar risk patterns, thereby improving the reliability of subsequent risk identification.

[0074] This invention can simultaneously utilize multiple types of user behavior data, such as registration information, login device, network environment, transaction behavior, browsing and clicking behavior, account association information, and historical risk control information, and adaptively enhance the effective information in different views. This avoids insufficient information utilization caused by insufficient single behavioral features or simple splicing, enabling the system to capture abnormal patterns of risky users from multiple angles and improve the ability to identify hidden risky accounts such as fraudulent users, users who exploit loopholes, malicious marketing users, and users who engage in bulk order placement.

[0075] In practical business applications, high-quality risk user tags often rely on manual review, historical penalty records, or post-event confirmation, which is costly to obtain. This invention can classify unlabeled users by leveraging user relationship structures and multi-view behavioral information, even with a small number of labeled users. This reduces the model's dependence on large-scale manually labeled data and is applicable to scenarios such as financial risk control, e-commerce anti-fraud, content community governance, and platform abnormal account identification. It can support practical business processing such as risk warning, manual review, and risk control interception. Attached Figure Description

[0076] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0077] Figure 1 This is a flowchart illustrating a specific embodiment of the present invention;

[0078] Figure 2 This is a schematic diagram of the framework of a specific embodiment of the present invention;

[0079] Figure 3 Line graph showing the accuracy of this invention and other methods at label rates of {5%,...50%} in six multi-view datasets;

[0080] Figure 4 Line graph showing the F1 scores of this invention and other methods at label rates of {5%,...50%} in six multi-view datasets;

[0081] Figure 5 This is a two-dimensional visualization of the invention on different datasets;

[0082] Figure 6These are the convergence curve, accuracy, and F1 score curves of the loss function of this invention. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Modifications to the present invention in various equivalent forms all fall within the scope defined by the appended claims.

[0085] like Figure 1 , 2 As shown, this invention provides a user risk classification method based on particle topology construction and feature enhancement, comprising the following steps:

[0086] Step 1: Acquire user behavior network data and construct it into multi-view user node data.

[0087] Each user account is treated as a user node, and the user's registration information, login device information, network environment information, transaction behavior information, browsing and clicking behavior information, account association information, and historical risk control information are respectively used as feature matrices under different views;

[0088] The multi-view user node data is as follows:

[0089] ;

[0090] In the formula, Number of user views; For the first User node feature representation under each view; For the first Number of feature dimensions per view;

[0091] No. The characteristics of user nodes under each view are represented as follows:

[0092] ;

[0093] In the formula, For the first The i-th column of features in each view has a row corresponding to a user node and a column corresponding to a user behavior feature.

[0094] Step 2: Perform unsupervised granular partitioning of user nodes under each view, and construct the initial user topology under the corresponding view based on the topology within and between granular nodes; learn the importance of edge connections between user nodes through a differentiable shrinking activation mechanism, and suppress noisy user relationship edges introduced by abnormal behavior masquerading, accidental similarity or weak correlation features to obtain a purified user relationship topology.

[0095] The specific process is as follows:

[0096] In topology construction, for the current user view, the N user nodes under the current view are set as the initial spheres, using... This serves as the stopping condition for particle-sphere splitting. First, k-means++ is used to set the sphere center. Then, the distance from the remaining user nodes to the sphere center is calculated, and the user nodes are assigned to sub-spheres. When the number of user nodes in a sphere is not greater than the stopping condition, the sphere stops splitting. This process is repeated until all spheres satisfy the condition, resulting in the th sphere. User node-level particle collection under each view:

[0097] ;

[0098] In the formula, GB represents the first... The spheres generated in the nth view; m is the nth view. The number of user node-level particles generated under each view;

[0099] Establish a connection between any two user nodes within the same sphere to generate the sphere's topology:

[0100] ;

[0101] In the formula, and The i-th and j-th user nodes are respectively in the i-th and j-th positions. Feature representation under each view; For the first The I-th particle in each view.

[0102] Calculate the representative node for each sphere; the representative user node is the mean of the feature representations of all user nodes within that sphere.

[0103] ;

[0104] In the formula, Let be the representative user node of the i-th sphere; This represents the number of user nodes in the current sphere.

[0105] Then, the Euclidean distance between user nodes represented by different spheres is calculated, and the nearest similar sphere is selected for each sphere:

[0106] ;

[0107] In the formula, The distance is Euclidean. Let be the representative user node of the i-th sphere; Let J be the representative user node of the J-th sphere;

[0108] Finally, user node connections are established between the particles to generate the inter-particle topology. :

[0109] ;

[0110] In the formula, and The i-th and j-th user nodes are respectively in the i-th and j-th positions. Feature representation under each view and The first The I-th and J-th particles under each view;

[0111] By fusing the intra-granular topology and inter-granular topology, we obtain the first... Initial topology in user view :

[0112] ;

[0113] In the formula, For the first Topology of the sphere under each view; For the first Topology of particles and spheres under each view;

[0114] By learning the importance of edge connections between user nodes through a differentiable shrinkable activation mechanism and suppressing noisy user relationship edges introduced by accidental similarity, anomalous behavior masquerading, or weakly correlated features, a purified user relationship topology is obtained:

[0115] ;

[0116] In the formula, The learnable edge coefficient matrix; The learnable threshold matrix; This is element-wise multiplication; It is a linear function; This is the initial user topology.

[0117] Step 3: For each user view, use the feature dimension as the granular sphere partitioning object, refine the features within the view, and concatenate the refined user feature representations of each view into a joint user feature;

[0118] The specific process is as follows:

[0119] For the current user view Using the feature dimensions in this view as the partitioning object, an unsupervised granular sphere method is employed to obtain multiple feature-level granular spheres:

[0120] ;

[0121] In the formula, For the first The set of feature-level particles under each user view; m is the number of feature-level particles;

[0122] To determine whether any two user feature dimensions belong to the same feature level sphere, if they do, calculate their cosine similarity; otherwise, set the corresponding similarity to zero. The feature similarity matrix is ​​constructed as follows:

[0123] ;

[0124] In the formula, For the first Feature similarity matrix of each user view; and The first The feature vectors corresponding to the i-th and j-th feature dimensions in each user view; Cosine similarity; For the first The i-th particle in a user view;

[0125] Based on the feature similarity matrix in step 32 Refine the original user characteristics under the current user view:

[0126] ;

[0127] In the formula, For the first Refined user feature representation under each user view;

[0128] By concatenating the refined feature representations of all user views, a joint feature representation is obtained:

[0129] ;

[0130] In the formula, For feature splicing; For user joint feature representation; For different view feature dimensions; N is the number of user nodes; R is the real number field.

[0131] Step 4: Generate a weight matrix using a multilayer perceptron based on the user joint features. Then, use the weight matrix to perform element-wise weighted enhancement on the original user feature representations of each view to obtain an enhanced user feature representation that integrates complementary information from the views.

[0132] The specific process is as follows:

[0133] Input the joint feature representation of the user into the view multilayer perceptron corresponding to each user's view to generate an adaptive weight matrix for each view:

[0134] ;

[0135] In the formula, For the first Weight matrix corresponding to each user view; For the first A multilayer perceptron corresponds to each user view; N is the number of user nodes; For the first Feature dimensions of each view;

[0136] The original user feature representation of the corresponding view is enhanced using a weight matrix:

[0137] ;

[0138] In the formula, For the first Enhanced feature representation under each user view; This is element-wise multiplication;

[0139] Step 5: Based on the purified user relationship topology, perform information propagation on the enhanced user feature representation to obtain the user node embedding representation under each view, and fuse the user node embedding representation of each view according to the learnable view fusion weight to obtain a unified user node representation;

[0140] The specific process is as follows:

[0141] The first Purified topology in user view and enhanced feature representation As input to a Graph Convolutional Network (GCN), the GCN adjusts the topology in each layer. The neighborhood range of user nodes is determined, and risk-related feature information is aggregated from neighboring user nodes to update the user node representation layer by layer. The information propagation process of the graph convolutional network is as follows:

[0142] ;

[0143] In the formula, For the first Node embedding under each user view, where... ; For activation functions; and The first The first view The learnable weight matrix of layer l and layer l;

[0144] The above process is performed on all user views. After passing through layer l of information, the user node embedding representations for each view are obtained. Learnable view fusion weights are set, and the user node embedding representations of each view are fused to obtain the final user node embeddings.

[0145] ;

[0146] In the formula, For the first The fusion weight of each user view, and satisfying .

[0147] Step 6: Represent the unified user nodes using a risk classifier to predict user risk, and identify normal users, low-risk users, medium-risk users, high-risk users, fraud risk users, or users at risk of exploiting promotional offers based on the predicted probabilities, in order to trigger risk warnings, manual review, or risk control interception actions.

[0148] The prediction probability matrix of the risk classifier is:

[0149] ;

[0150] In the formula, H represents the embedding representation of the user node;

[0151] Based on the actual labels and predicted probability matrix of the labeled user nodes, calculate the classification loss:

[0152] ;

[0153] In the formula, The set of labeled user nodes; Let i be the true label of the i-th user node belonging to the c-th class; This represents the probability that the i-th user node is predicted to be of class c.

[0154] The effects of the present invention will be illustrated by simulation below.

[0155] Experimental dataset and experimental setup:

[0156] To verify the effectiveness and generalization ability of this invention in multi-view user node classification scenarios, this embodiment uses internationally recognized multi-view standard datasets for equivalent simulation testing. In the equivalent simulation, each sample is considered as a user node to be classified, and different view features are used to simulate user behavior information under different business sources, such as registration information, device environment, transaction behavior, browsing and clicking behavior, and account association behavior. This embodiment uses six multi-view datasets for experiments: ALOI, Caltech101-7, GRAZ02, NUS-WIDE, OutScene, and Scene15. Table 1 below provides basic information about the six datasets, including the number of samples, the number of views, the feature dimensions of each view, the number of categories, and the data type.

[0157] Table 1. Statistics of 6 Multi-view Image Datasets

[0158]

[0159] To verify the advantages of this invention over traditional methods, this embodiment selects several multi-view semi-supervised node classification methods as comparison methods, including: Co-GCN, DSRL, LGCN-FF, JFGCN, IMvGCN, GEGCN, ICGCN, and GBCM-GCN.

[0160] Regarding dataset partitioning, this embodiment employs a unified data partitioning strategy: randomly selecting 10% of the samples in each dataset as the training set, 10% as the validation set, and the remaining 80% as the test set. The training set is used for model parameter optimization, the validation set for parameter selection, and the test set for evaluating the final classification performance.

[0161] To further simulate the scenario of limited labels in actual user risk control, this embodiment also sets up experiments under different label ratios. Specifically, the training sample ratio is gradually increased from 5% to 50% in 5% increments to observe the classification performance of the method of the present invention under different label scarcity levels. This setting can simulate the business scenario in a real platform where only a small number of confirmed risk user labels are available, while risk identification of a large number of unlabeled users is required. Experimental results show that the method of the present invention can still maintain superior performance under low label ratio conditions.

[0162] To comprehensively evaluate the user node risk classification performance of the method of this invention, this embodiment uses accuracy (ACC) and F1 score as evaluation metrics. Accuracy measures the overall classification correctness of the model, while F1 score comprehensively measures precision and recall, and is more suitable for evaluating the overall performance in multi-class node classification tasks. Experimental results also use ACC and F1 score to compare the performance of different methods.

[0163] This embodiment is implemented based on the PyTorch deep learning framework and runs on a computer equipped with an Intel i9-12400F CPU, an NVIDIA GeForce RTX 4070 GPU, and 32 GB of RAM. For model training settings, this embodiment uses a two-layer graph convolutional structure, with ReLU activation function in the hidden layer and Softmax activation function in the output layer. During training, the maximum number of training epochs is set to 300, the learning rate is set to 0.01, the optimizer is Adam, and the weight decay coefficient is set to... .

[0164] Comparative experimental results:

[0165] After end-to-end training and testing, the statistical results of ACC and F1-score for each model on the multi-view node classification task are shown in Table 2.

[0166] Table 2. Performance Comparison of MGCN-CE and 8 Multi-View Learning Methods in Semi-Supervised Risk Identification

[0167]

[0168] The experimental data in Table 2 fully demonstrate that the MGCN-CE method of this invention has significant improvements and beneficial effects compared with the traditional multi-view graph convolution method.

[0169] MGCN-CE achieved the highest ACC and F1 score on all six datasets. On the Caltech101-7, GRAZ02, NUS-WIDE, OutScene, and Scene15 datasets, MGCN-CE improved the ACC by 1.4%, 4.3%, 3.1%, 5.6%, and 1.8% respectively compared to the suboptimal GBCM-GCN method. This performance improvement is mainly attributed to the topology building module utilizing both intra- and inter-particle connections. Compared to GBCM-GCN, which only uses inter-particle connections, the proposed topology building method generates a more reliable topology structure, thus contributing to higher-quality node embedding representations.

[0170] Compared to multi-view methods that primarily focus on inter-view consistency, such as LGCN-FF, JFGCN, IMvGCN, ICGCN, and GBCM-GCN, MGCN-CE further mines complementary information between views and effectively enhances feature representations, thus achieving better classification performance. Furthermore, compared to methods that do not simultaneously consider topology construction and feature enhancement, such as Co-GCN, IMvGCN, GEGCN, and ICGCN, MGCN-CE achieves superior classification performance on all datasets by jointly optimizing topology structure and feature representations.

[0171] Contributions to the core methodology:

[0172] To verify the effectiveness of the core methods in this invention, this embodiment constructs an ablation model based on the complete model MGCN-CE and compares it with the complete model. Specifically, two ablation variants are set: First, the particle-sphere topology construction module in this invention is replaced with the traditional kNN graph construction method to obtain the model. Second, the feature enhancement module in this invention is removed, and the original features are directly input into the subsequent graph convolutional network to obtain the model. The contributions of the particle-sphere topology construction module and the feature enhancement module to the final classification performance can be verified using the two ablation models described above. The experimental results are shown in Table 3.

[0173] Table 3 Ablation experimental results of MGCN-CE on different datasets

[0174]

[0175] As shown in Table 3, the complete model MGCN-CE outperforms the two ablation models in both ACC and F1-score across all datasets, indicating that the topology building module and feature enhancement module in this invention play an important role in improving model performance.

[0176] Secondly, after removing the feature enhancement module, the model The performance was also lower than the full model on all datasets, with the most significant drop in accuracy (3.6%) on the ALOI dataset. This demonstrates that the feature enhancement module of this invention can effectively utilize complementary information between multiple views to improve the quality of user node feature representations, thereby improving the final classification performance.

[0177] In summary, the ablation experiments demonstrate that the performance improvement of this invention does not stem from a single module, but rather from the combined effect of the particle-sphere topology construction module and the multi-view feature enhancement module. Specifically, the topology construction module primarily improves the reliability of the user relationship structure and reduces the interference of noisy connections on information propagation; the feature enhancement module primarily enhances the multi-view user feature representation capability, enabling full utilization of complementary information between different views. Therefore, this invention can achieve more stable and accurate classification results in user risk classification scenarios.

[0178] Robustness analysis under tag scarcity conditions:

[0179] To verify the classification performance of the method under different labeling conditions, this embodiment further sets up experiments with different labeling rates. The labeling rate represents the proportion of training samples to the total number of samples. In the experiment, the proportion of training samples is gradually increased from 5% to 50% in a step size of 5%, and the ACC and F1-score of different methods under each labeling rate condition are recorded respectively. Figure 3 and Figure 4 The changes in ACC and F1-score for different methods under different labeling rates are shown respectively.

[0180] Experimental results show that the classification performance of each method generally increases with the increase of the labeling rate, indicating that more labeled samples can provide more sufficient supervision information for model training. Meanwhile, the proposed method MGCN-CE maintains superior performance under different labeling rate conditions, and its advantages are more pronounced under low labeling rate conditions.

[0181] The results demonstrate that, even with a limited number of labeled samples, the method of this invention can utilize not only the supervisory information provided by labeled user nodes but also the multi-view behavioral features of unlabeled user nodes and their similarity relationships with other user nodes for representation learning. Therefore, in real-world user risk classification scenarios, even if the platform only possesses a small number of confirmed risky or normal user samples, this invention can still achieve relatively stable risk identification results.

[0182] Visual analysis:

[0183] like Figure 5 As shown, compared to GEGCN and ICGCN, the node embeddings learned by the method of this invention, MGCN-CE, exhibit more pronounced intra-class clustering and inter-class separation. That is, nodes of the same class are more concentrated in the two-dimensional space, and the overlapping areas between different classes are significantly reduced. Although a small number of outliers still exist, the node representation generated by the method of this invention has a clearer overall distribution, and the outlier dispersion is lower than that of the comparative methods.

[0184] Convergence analysis:

[0185] To verify the stability of the method of the present invention during the training process, this embodiment further analyzes the convergence of the model. For example... Figure 6 As shown, within the first 100 epochs of training, the loss function decreases rapidly, while the ACC and F1-score increase significantly, indicating that the model can complete effective learning in a relatively small number of iterations. As the training process continues, the model gradually enters the convergence phase, the loss function tends to stabilize, and the ACC and F1-score remain relatively stable, indicating that the model has been sufficiently optimized and has good training stability.

[0186] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A user risk classification method based on granular sphere topology construction and feature enhancement, characterized in that, Includes the following steps: Step 1: Obtain user behavior network data and construct it into multi-view user node data, where each user account is a user node and the network data of each user are feature matrices under different views; Step 2: Perform unsupervised granular partitioning on user nodes under each view, and construct the initial user topology under the corresponding view based on the topology within and between granular spheres; suppress noisy user relationship edges through a differentiable shrinking activation mechanism to obtain a purified user relationship topology; Step 3: Refine the features within each view by using feature-level granular sphere partitioning, and then concatenate the refined user feature representations of each view into a user joint feature; Step 4: Generate a weight matrix using a multilayer perceptron based on the user joint features. Then, use the weight matrix to perform element-wise weighted enhancement on the original user feature representations of each view to obtain an enhanced user feature representation that integrates complementary information from the views. Step 5: Based on the purified user relationship topology, perform information propagation on the enhanced user feature representation to obtain the user node embedding representation under each view, and fuse the user node embedding representation of each view according to the learnable view fusion weight to obtain a unified user node representation; Step 6: Use a risk classifier to predict user risk by representing the unified user nodes, and obtain user risk classification.

2. The user risk categorization method based on granule topology construction and feature enhancement as claimed in claim 1, wherein, In step 2, the method for obtaining the topology within the spheres during the unsupervised sphere partitioning process is as follows: Set the N user nodes in the current view as the initial spheres, so that... As a stopping condition for particle splitting, k-means++ is used to set the sphere center and divide it into sub-spheres, obtaining a set of user node-level particles. A connection is established between any two user nodes within the same particle to generate the intra-sphere topology. The specific formula for the user node-level particle set is as follows: ； In the formula, GB represents the first... The spheres generated in the nth view; m is the nth view. The number of user node-level particles generated under each view; The internal topology of the generated spheres is as follows: ； In the formula, and The i-th and j-th user nodes are respectively in the i-th and j-th positions. Feature representation under each view; For the first The I-th particle in each view.

3. The user risk classification method based on particle topology construction and feature enhancement as described in claim 1 or 2, characterized in that, In step 2, the method for obtaining the topology between particles during the unsupervised particle partitioning process is as follows: Calculate the representative node for each sphere, determine similar spheres based on the Euclidean distance between the representative nodes of different spheres, and generate the sphere topology by establishing user node connections between similar spheres. The representative user node is the mean of the feature representations of all user nodes within the current sphere, and the specific formula is as follows: ； In the formula, Let be the representative user node of the i-th sphere; This represents the number of user nodes in the current sphere. The formula for calculating the Euclidean distance between nodes represented by different spheres is as follows: ； In the formula, The distance is Euclidean. Let be the representative user node of the i-th sphere; Let J be the representative user node of the J-th sphere; The resulting intergranular topology is as follows: ； In the formula, and The i-th and j-th user nodes are respectively in the i-th and j-th positions. Feature representation under each view and The first The I-th and J-th particles under each view.

4. The user risk classification method based on particle topology construction and feature enhancement as described in claim 3, characterized in that, In step 2, the method for constructing the initial user topology is as follows: The intra-sphere topology and inter-sphere topology are merged to obtain the initial user topology, as shown in the following formula: ； In the formula, For the first Topology of the sphere under each view; For the first The topology between particles and spheres under each view.

5. The user risk classification method based on particle topology construction and feature enhancement as described in claim 1, characterized in that, In step 2, the specific method for suppressing noisy user relationship edges and obtaining a purified user relationship topology through a differentiable shrinkable activation mechanism is as follows: The importance of edge connections between user nodes in the initial user topology is learned through a differentiable shrinkable activation mechanism, and noisy user relationship edges introduced by accidental similarity, anomalous behavior masquerading, and / or weak correlation features are suppressed. The specific formula is as follows: ； In the formula, The learnable edge coefficient matrix; The learnable threshold matrix; This is element-wise multiplication; It is a linear function; This is the initial user topology.

6. The user risk classification method based on particle topology construction and feature enhancement as described in claim 1, characterized in that, In step 3, the specific method for segmenting characteristic-level particles is as follows: For each user view, the feature dimensions in that view are used as the partitioning object for unsupervised granular sphere partitioning, resulting in multiple feature-level granular spheres. The specific formula is as follows: ； In the formula, For the first The set of feature-level particles under each user view; m is the number of feature-level particles.

7. The user risk classification method based on particle topology construction and feature enhancement as described in claim 1 or 6, characterized in that, In step 3, the specific method for refining features within the view is as follows: Construct a feature similarity matrix; The original user feature representation in the current view is refined within the view using a feature similarity matrix to obtain a refined user feature representation. The specific method for constructing the feature similarity matrix is ​​as follows: To determine whether any two user feature dimensions belong to the same feature level sphere, if they do, calculate their cosine similarity; if they do not, set the corresponding similarity to zero. The specific formula for constructing the feature similarity matrix is ​​as follows: ； In the formula, For the first Feature similarity matrix of each user view; and The first The feature vectors corresponding to the i-th and j-th feature dimensions in each user view; Cosine similarity; For the first The i-th particle in a user view; The specific formula for refining features within a view is as follows: ； In the formula, For the first Refined user feature representation under each user view; The specific formula for user joint features is: ； In the formula, For feature splicing; For user joint feature representation; represents the feature dimensions of different views; N is the number of user nodes; R is the real number field.

8. The user risk classification method based on particle topology construction and feature enhancement as described in claim 1, characterized in that, In step 4, the specific formula for obtaining the weight matrix is ​​as follows: ； In the formula, For the first Weight matrix corresponding to each user view; For the first A multilayer perceptron corresponds to each user view; N is the number of user nodes; For the first Feature dimensions of each view; The specific formula for obtaining the enhanced user feature representation is as follows: ； In the formula, For the first Enhanced feature representation under each user view; This is element-wise multiplication.

9. The user risk classification method based on particle topology construction and feature enhancement as described in claim 1, characterized in that, In step 5, the specific method for performing information dissemination is as follows: The purified user relationship topology and enhanced user feature representation under each view are used as input to the graph convolutional network. Based on the purified user relationship topology, the neighborhood range of user nodes is determined, and the feature information of neighboring user nodes is aggregated to obtain the user node embedding representation under the corresponding view. The specific formula for the user node embedding representation is as follows: ； In the formula, ; For the first Node embedding under each user view, where... ; For activation functions; and The first The first view The learnable weight matrix of layer l and layer l; For the first The fusion weight of each user view.

10. The user risk classification method based on particle topology construction and feature enhancement as described in claim 1, characterized in that, In step 6, the prediction probability matrix of the risk classifier is: ； In the formula, H represents the embedding representation of the user node; The classification loss is: ； In the formula, The set of labeled user nodes; Is the i-th user node a true label belonging to the c-th class? This represents the probability that the i-th user node is predicted to be of class c.

Images