Transaction platform node attribute information detection model training method and training device

Abstract

The present application relates to the technical field of artificial intelligence, and more particularly to a transaction platform node attribute information detection model training method and training device. The model training method comprises: obtaining graph network data of a transaction platform, selecting part of the nodes in the graph network data as sample nodes, and labeling the true attribute information of each sample node; the attribute information of the nodes includes abnormal nodes or normal nodes; inputting the graph network data into an initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node; and updating the model parameters of the initial node attribute information detection model according to the true attribute information and the predicted attribute information of the sample nodes to obtain a trained node attribute information detection model. The node attribute information detection model obtained through deep learning can more accurately detect the node attribute information in the transaction platform.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a training method and training device for a node attribute information detection model in a trading platform. Background Technology

[0002] With the development of internet technology, online shopping and online transactions have brought convenience to people's lives, leading to the emergence of various online trading platforms.

[0003] Because the identities of traders on online trading platforms are often uncertain, fraudulent transactions exist. For example, some fraudulent users impersonate legitimate users to conduct transactions. These users mimic legitimate users to conceal their true identities and contact legitimate users more frequently, making it difficult for traders to distinguish between them. Therefore, it is necessary to detect these fraudulent users on online trading platforms for further processing. Current technologies for user identity verification on online trading platforms typically treat the platform as a graph network, with users as nodes, and node attributes representing user attributes.

[0004] In existing technologies, when detecting abnormal users on a trading platform, big data analysis combined with manual analysis is usually used to determine the attribute information of a certain node on the online trading platform. However, the method of using big data analysis combined with manual analysis still has technical problems of low detection efficiency and low detection accuracy. Summary of the Invention

[0005] This invention provides a training method and apparatus for a node attribute information detection model in a trading platform, in order to solve the technical problem of low detection accuracy in the prior art for detecting the existence of node attributes in online trading platforms.

[0006] On one hand, the present invention provides a training method for a node attribute information detection model in a trading platform, comprising:

[0007] Obtain graph network data from the trading platform;

[0008] A subset of nodes in the graph network data are selected as sample nodes, and the true attribute information of each sample node is labeled; the attribute information of the node includes abnormal nodes or normal nodes;

[0009] The graph network data is input into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node;

[0010] The model parameters of the initial node attribute information detection model are updated based on the actual attribute information and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model.

[0011] According to the present invention, a training method for a node attribute information detection model in a trading platform is provided, wherein the initial node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network;

[0012] The step of inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node includes:

[0013] The graph network data is input into the decoupled network to obtain the topology view data and attribute view data corresponding to the graph network data;

[0014] The topology view data is input into the first multilayer perceptron network, and the attribute view data is input into the second multilayer perceptron network. The first multilayer perceptron network and the second multilayer perceptron network are respectively used to perform dimensionality reduction processing on the topology view data and the attribute view data to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data.

[0015] The topological feature map and attribute feature map are input into the feature fusion network to obtain the fused feature information;

[0016] The fused feature information is input into the linear classification network model to obtain the predicted attribute information corresponding to each sample node.

[0017] According to the training method of the node attribute information detection model in the trading platform provided by the present invention, the initial node attribute information detection model further includes: a first maximization processing module, a second maximization processing module, and a minimization processing module;

[0018] The step of inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node further includes:

[0019] The topology view data and topology feature map are input into the first maximization processing module to obtain the corresponding topology feature information; the attribute view data and attribute feature map are input into the second maximization processing module to obtain the corresponding attribute feature information.

[0020] The topological feature map and attribute feature map are input into the minimization processing module to obtain fused semantic information;

[0021] The topological feature information, attribute feature information, and fused semantic information are input into the feature fusion network to obtain the fused feature information.

[0022] According to the training method of the node attribute information detection model in the trading platform provided by the present invention, the loss function of the initial node attribute information detection model is as follows (1):

[0023]

[0024] in,

[0025] In formula (2), v i This represents the i-th sample node. Let y represent the set of sample nodes. i Let z represent the true label of the i-th node, σ represent the non-linear activation function, W′ represent the weight matrix of the linear classification network, and z i b represents the feature learned from the i-th node, and b′ represents the bias vector of the linear classification network.

[0026]

[0027]

[0028] In formulas (3) and (4), x1 represents the index of the topological feature information, x2 represents the index of the attribute feature information, z1 represents the topological feature map, and z2 represents the attribute feature map; q(x i |z i ) represents x i and z i Variational estimation under conditional distributions, where p(z1, x2) represents the joint probability distribution of z1 and x2, E represents the expectation, r(z1) represents the marginal distribution of z1, and r(z2) represents the marginal distribution of z2. Let z2 represent the probability distribution of z2 under the condition x2 in the feature space of x2; This indicates that in the feature space of x1, z1 lies in x. 21 The probability distribution under the given conditions.

[0029] On the other hand, the present invention also provides a method for detecting node attribute information in a trading platform, comprising:

[0030] Obtain the graph network data of the trading platform to be tested;

[0031] The graph network data of the transaction platform to be tested is input into the node attribute information detection model to obtain the attribute information corresponding to each node in the transaction platform to be tested; the attribute information includes abnormal nodes or normal nodes.

[0032] The node attribute information detection model is any one of the node attribute information detection models described above.

[0033] According to the present invention, a method for detecting node attribute information in a trading platform is provided, wherein the node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network;

[0034] The step of inputting the graph network data of the transaction platform to be tested into the node attribute information detection model to obtain the attribute information corresponding to each node in the transaction platform to be tested includes:

[0035] The graph network data of the transaction platform to be tested is input into the decoupled network to obtain the topology view data and attribute view data corresponding to the transaction platform to be tested.

[0036] The topology view data and attribute view data corresponding to the transaction platform to be detected are respectively input into the first multilayer perceptron network and the second multilayer perceptron network;

[0037] The first multilayer perceptron network and the second multilayer perceptron network are respectively used to perform dimensionality reduction processing on the topology view data and the attribute view data to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data.

[0038] The topological feature map and attribute feature map are input into the feature fusion network to obtain the fused feature information corresponding to the transaction platform to be detected.

[0039] The fused feature information corresponding to the transaction platform to be detected is input into the linear classification network model to obtain the attribute information corresponding to each node in the transaction platform to be detected.

[0040] According to the present invention, a method for detecting node attribute information in a trading platform, the node attribute information detection model further includes: a first maximization processing module, a second maximization processing module, and a minimization processing module;

[0041] The step of inputting the graph network data of the transaction platform to be tested into the node attribute information detection model to obtain the attribute information corresponding to each node in the transaction platform to be tested also includes:

[0042] The topology view data and topology feature map corresponding to the transaction platform to be detected are input into the first maximization processing module to obtain the corresponding topology feature information; the attribute view data and attribute feature map are input into the second maximization processing module to obtain the corresponding attribute feature information.

[0043] The topological feature map and attribute feature map are input into the minimization processing module to obtain fused semantic information;

[0044] The topological feature information, attribute feature information, and fused semantic information are input into the feature fusion network to obtain the fused feature information corresponding to the transaction platform to be detected.

[0045] According to a method for detecting node attribute information in a trading platform provided by the present invention, the detection method further includes:

[0046] Construct the graph network corresponding to the transaction platform to be tested;

[0047] Feature extraction is performed on the graph network corresponding to the transaction platform to be tested to obtain the graph network data of the transaction platform to be tested.

[0048] The present invention also provides a training device for a node attribute information detection model in a trading platform, comprising:

[0049] The acquisition unit is used to acquire graph network data from the trading platform;

[0050] The sample annotation unit is used to select a portion of the nodes in the graph network data as sample nodes and to annotate the true attribute information of each sample node; the attribute information of the node includes abnormal nodes or normal nodes.

[0051] The processing unit is used to input the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node;

[0052] The update unit is used to update the model parameters of the initial node attribute information detection model according to the real attribute information and predicted attribute information of the sample nodes, so as to obtain the trained node attribute information detection model.

[0053] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a training method for a node attribute information detection model in a trading platform as described in any of the preceding claims; or implements a node attribute information detection method in a trading platform as described in any of the preceding claims.

[0054] The present invention provides a training method for a node attribute information detection model in a trading platform. This method involves acquiring graph network data from the trading platform, selecting a subset of nodes as sample nodes, and labeling the true attribute information of each sample node. The node attribute information includes both abnormal and normal nodes. The graph network data is then input into an initial node attribute information detection model to obtain the predicted attribute information for each sample node. The model parameters of the initial node attribute information detection model are updated based on the true and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model. This node attribute information detection model, obtained through deep learning, can more accurately detect node attribute information in the trading platform. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0056] Figure 1 A flowchart illustrating the training method for a node attribute information detection model in a trading platform provided in this embodiment of the invention;

[0057] Figure 2 This is a schematic diagram of the structure of the initial node attribute information detection model according to an embodiment of the present invention;

[0058] Figure 3 This is a signal flow diagram in the initial node attribute information detection model of this invention embodiment;

[0059] Figure 4 This is a performance comparison chart of the node attribute information detection model of this invention and existing models;

[0060] Figure 5 This invention provides a method for detecting node attribute information in a trading platform.

[0061] Figure 6 A schematic diagram of the structure of a training device for a node attribute information detection model in a trading platform provided in an embodiment of the present invention;

[0062] Figure 7 This is a schematic diagram of the structure of a node attribute information detection device in a trading platform provided by an embodiment of the present invention;

[0063] Figure 8 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0065] Graph-based anomaly detection is a crucial task with significant applications, such as anomaly detection of opinions, fake news detection, spam detection, and financial fraud detection. In these scenarios, graph structures effectively model the relationships between entities. In this invention, interactive activities on a trading platform can be viewed as a graph, with users or items often represented as nodes and transactions or relationships as edges. Currently, numerous techniques exist for detecting anomalous users.

[0066] The training method for the node attribute information detection model in the trading platform provided by this invention will be described in detail below through several specific embodiments. It is understood that these specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0067] like Figure 1 This is a flowchart illustrating a training method for a node attribute information detection model in a trading platform provided in an embodiment of the present invention. This training method can be executed by software and / or hardware devices. For example, the hardware device can be an embedded device, a personal computer, or a server, etc. For further examples, please refer to... Figure 1 As shown, the training method for this node attribute information detection model may include:

[0068] S101. Obtain graph network data from the trading platform.

[0069] For example, the interactive activities on a trading platform can be viewed as a graph, with users or items often treated as nodes and transactions or relationships often treated as edges, thus obtaining the graph network data of the trading platform.

[0070] S102. Select some nodes in the graph network data as sample nodes and label the true attribute information of each sample node; the attribute information of the node includes abnormal nodes or normal nodes.

[0071] Since the information of each node in the graph is related to its neighboring nodes, the detection of each node requires understanding its relationship with its neighbors. Therefore, in this embodiment of the invention, it is not possible to directly obtain the relevant data of the sample nodes. Furthermore, forcibly segmenting the data of the sample nodes in this way would result in the loss of some information from some sample nodes, affecting the detection results. Therefore, in this embodiment, a subset of nodes in the graph network data are directly selected as sample nodes, and the true attribute information of each sample node is labeled. The labeled information for the sample nodes includes abnormal nodes or normal nodes.

[0072] S103. Input the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node.

[0073] Since the data information of the sample nodes (i.e., the training data) cannot be directly segmented, in this embodiment, the entire graph network data is directly input into the initial node attribute information detection model.

[0074] Due to the powerful representational capabilities of graph structures and the rapid development of Graph Neural Networks (GNNs), many methods have attempted to use GNNs for anomaly detection on homogeneous or heterogeneous graphs. Their main idea is to utilize GNNs to learn expressive node representations, distinguishing normal and anomaly nodes in the latent space. Message-passing GNNs have been a mainstream approach in recent years, achieving local smoothing through neighbor aggregation and layer stacking. While message-passing GNNs can detect anomalies in most scenarios, the strong inductive bias of homoophily hinders their performance on heteroophilic graphs. Most current GNNs can be viewed as low-pass filters, resulting in poor generalization when the graph signal is high-frequency. In anomaly detection tasks, anomaly nodes often mimic normal nodes to conceal themselves, interacting more frequently with them. For example, in the YelpChi dataset, 81% of the neighbors of anomaly nodes are normal nodes. Therefore, traditional message-passing GNNs cannot address the inconsistency problem in anomaly detection graph datasets.

[0075] like Figure 2 This is a schematic diagram of the structure of the initial node attribute information detection model according to an embodiment of the present invention. The initial node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network.

[0076] For example, such as Figure 3This is the signal flow graph in the initial node attribute information detection model of this embodiment of the invention. The graph network data is input into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node, including:

[0077] The graph network data is input into a decoupled network to obtain the topology view data (i.e., adjacency matrix) and attribute view data (i.e., attribute matrix) corresponding to the graph network data. The topology view data is input into a first multilayer perceptron network, and the attribute view data is input into a second multilayer perceptron network. The first and second multilayer perceptron networks are used to perform dimensionality reduction processing on the topology view data and attribute view data, respectively, to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data.

[0078] For example, this embodiment uses two multilayer perceptron networks as encoders to obtain feature maps corresponding to topology view data and attribute view data.

[0079] The topological feature map and attribute feature map are input into a feature fusion network to obtain fused feature information. Preferably, the feature fusion network in this embodiment is an attention mechanism network, which dynamically fuses the features of the topological view data and attribute view data using an attention mechanism to obtain a unified learned representation. The two attention coefficients represent the weights of the topological feature map and the attribute feature map, respectively, and are then normalized using a softmax function. A larger weight coefficient indicates that the corresponding feature map is more important. The final representation can be obtained by multiplying the representations of the topological feature map and the attribute feature map by their corresponding attention coefficients and then adding them together.

[0080] Finally, the fused feature information is input into the linear classification network model to obtain the predicted attribute information corresponding to each sample node.

[0081] In another embodiment, the initial node attribute information detection model further includes: a first maximization processing module, a second maximization processing module, and a minimization processing module.

[0082] For example, inputting graph network data into an initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node also includes:

[0083] To further decouple the information from the topology feature map and the attribute feature map, this embodiment maximizes the mutual information between the feature map and the input view data to ensure that semantic information is maintained while decoupling. For example, the topology view data and the topology feature map are input into the first maximization processing module to obtain the corresponding topology feature information; the attribute view data and the attribute feature map are input into the second maximization processing module to obtain the corresponding attribute feature information.

[0084] To further reduce the conflict between node attribute features and topological features, this embodiment of the invention uses the theory of variational information bottleneck to minimize the mutual information between the two. This not only reduces the conflict between the two types of features but also reduces redundant information. For example, the topological feature map and attribute feature map are input into the minimization processing module to obtain fused semantic information. Finally, the topological feature information, attribute feature information, and fused semantic information are input into the feature fusion network to obtain fused feature information. The fused feature information is then input into a linear classification network model to obtain the predicted attribute information corresponding to each sample node.

[0085] S104. Update the model parameters of the initial node attribute information detection model based on the real attribute information and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model.

[0086] For example, when updating the model parameters of the initial node attribute detection model based on the real attribute information and predicted attribute information of the sample nodes, the model parameters are continuously updated according to the model's loss function until the updated initial node attribute detection model converges, and the converged initial node attribute detection model is determined as the final trained node attribute detection model.

[0087] For example, the loss function of the initial node attribute information detection model in this embodiment of the invention is the following formula (1):

[0088]

[0089] in,

[0090] In formula (2), v i This represents the i-th sample node. Let y represent the set of sample nodes. i Let z represent the true label of the i-th node, σ represent the non-linear activation function, W′ represent the weight matrix of the linear classification network, and z i Let b' represent the feature learned from the i-th node, and b' represent the bias vector of the linear classification network.

[0091]

[0092]

[0093] In formulas (3) and (4), x1 represents the index of topological feature information, x2 represents the index of attribute feature information, z1 represents the topological feature map, and z2 represents the attribute feature map; q(x i |z i ) represents x i and z iVariational estimation under conditional distributions, where p(z1,x2) represents the joint probability distribution of z1 and x2, E represents the expectation, r(z1) represents the marginal distribution of z1, and r(z2) represents the marginal distribution of z2. Let z2 represent the probability distribution of z2 under the condition x2 in the feature space of x2; This indicates that in the feature space of x1, z1 lies in x. 21 The probability distribution under the given conditions.

[0094] For example, in order to make the detection of node attribute information in the trading platform more accurate, the sample nodes used to train the initial node attribute information detection model are selected from some nodes in the trading platform to be identified. In this way, the node attribute information detection model trained will be more accurate in detecting the node attribute information in the trading platform to be monitored.

[0095] For example, after the initial node attribute information detection model is trained, a portion of the verification nodes (i.e., verification data) are usually selected to detect the attribute information of the verification nodes in order to verify the accuracy of the node attribute information detection model trained.

[0096] For example, in a graph network dataset of a trading platform to be identified, all nodes are typically allocated as follows: 40% for training, 20% for validation, and 40% for testing. The test set consists of the nodes to be tested. To mitigate class imbalance in the training set, this method uses downsampling or reweighting during training.

[0097] For example, in this embodiment, YelpChi and Amazon, two common real-world datasets, are selected as the datasets for evaluating the performance of the node attribute information detection model. Figure 4 To verify the effectiveness of the proposed model, this embodiment selects several representative models for comparative experiments with the proposed model, including GCN, GAT, GraphSAGE, DR-GCN, CARE-GNN, FRAUDRE, and PC-GNN.

[0098] Since the samples in the test set are also class-imbaled, accuracy cannot accurately assess the effectiveness of the model. Therefore, this embodiment uses F1-macro, AUC, and GMean as evaluation metrics. Please refer to [link / reference]. Figure 4The method in this embodiment significantly outperforms the original model across all metrics on the YelpChi and Amazon datasets. While existing methods have achieved a high level of performance on the Amazon dataset with limited improvement, the model in this embodiment still demonstrates a significant performance improvement. The comparison methods in this embodiment can be broadly categorized into two types: traditional graph neural networks based on message passing mechanisms, such as GCN, GAT, and GraphSAGE; and the graph-based node attribute information detection model used in this embodiment. Because the node attribute information detection model considers inconsistencies in abnormal networks, it significantly outperforms traditional message-passing-based graph neural networks on the YelpChi and Amazon datasets, resulting in higher accuracy in identifying node attributes.

[0099] As can be seen, in this embodiment of the invention, when training the node attribute information detection model, the graph network data of the trading platform to be identified can be obtained first. A portion of the nodes in the graph network data are selected as sample nodes, and the true attribute information of each sample node is labeled. The attribute information of the nodes includes abnormal nodes or normal nodes. The graph network data is input into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node. The model parameters of the initial node attribute information detection model are updated based on the true and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model. Thus, the node attribute information detection model obtained through deep learning can more accurately detect the node attribute information in the trading platform.

[0100] like Figure 5 The present invention provides a method for detecting node attribute information in a trading platform, the method comprising:

[0101] S501: Obtain the graph network data of the trading platform to be tested.

[0102] For example, firstly, a graph network corresponding to the trading platform to be detected is constructed; then, features are extracted from the graph network corresponding to the trading platform to be detected to obtain the graph network data of the trading platform to be detected.

[0103] S502: Input the graph network data of the trading platform to be tested into the node attribute information detection model to obtain the attribute information corresponding to each node in the trading platform to be tested.

[0104] The attribute information includes abnormal nodes or normal nodes; the node attribute information detection model is the node attribute information detection model shown in the above embodiment.

[0105] For example, such as Figure 2The node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network.

[0106] For example, the graph network data of the trading platform to be tested is input into the node attribute information detection model to obtain the attribute information corresponding to each node in the trading platform to be tested, including:

[0107] The graph network data of the trading platform to be tested is input into a decoupled network to obtain the topology view data and attribute view data corresponding to the trading platform. The topology view data and attribute view data are then input into a first-layer perceptron network and a second-layer perceptron network, respectively. The first-layer perceptron network and the second-layer perceptron network are used to perform dimensionality reduction processing on the topology view data and attribute view data, respectively, to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data. The topology feature map and attribute feature map are then input into a feature fusion network to obtain the fused feature information corresponding to the trading platform to be tested. Finally, the fused feature information corresponding to the trading platform to be tested is input into a linear classification network model to obtain the attribute information corresponding to each node in the trading platform to be tested.

[0108] For example, the node attribute information detection model also includes: a first maximization processing module, a second maximization processing module, and a minimization processing module.

[0109] For example, the graph network data of the trading platform to be tested is input into the node attribute information detection model to obtain the attribute information corresponding to each node in the trading platform to be tested, which also includes:

[0110] The topology view data and topology feature map corresponding to the transaction platform to be detected are input into the first maximization processing module to obtain the corresponding topology feature information; the attribute view data and attribute feature map are input into the second maximization processing module to obtain the corresponding attribute feature information; the topology feature map and attribute feature map are input into the minimization processing module to obtain fused semantic information; the topology feature information, attribute feature information and fused semantic information are input into the feature fusion network to obtain the fused feature information corresponding to the transaction platform to be detected.

[0111] This invention uses a node attribute information detection model to predict the node attribute information of a trading platform, making the detection of node attribute information in the trading platform more accurate.

[0112] like Figure 6 This is a schematic diagram of the structure of a training device 60 for a node attribute information detection model in a trading platform provided in an embodiment of the present invention. The training device includes: an acquisition unit 601, a sample labeling unit 602, a processing unit 603, and an update unit 604.

[0113] The system includes the following components: an acquisition unit 601, which acquires graph network data from the trading platform; a sample labeling unit 602, which selects a subset of nodes from the graph network data as sample nodes and labels the true attribute information of each sample node, including whether the node is an anomaly or a normal node; a processing unit 603, which inputs the graph network data into an initial node attribute information detection model to obtain the predicted attribute information for each sample node; and an update unit 604, which updates the model parameters of the initial node attribute information detection model based on the true and predicted attribute information of the sample nodes to obtain a trained node attribute information detection model.

[0114] Optionally, the initial node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network.

[0115] Processing unit 603 is specifically used to input graph network data into a decoupled network to obtain topology view data and attribute view data corresponding to the graph network data; input the topology view data into a first multilayer perceptron network and the attribute view data into a second multilayer perceptron network. The first and second multilayer perceptron networks are used to perform dimensionality reduction processing on the topology view data and attribute view data, respectively, to obtain topology feature maps corresponding to the topology view data and attribute feature maps corresponding to the attribute view data. The topology feature maps and attribute feature maps are then input into a feature fusion network to obtain fused feature information. Finally, the fused feature information is input into a linear classification network model to obtain predicted attribute information corresponding to each sample node.

[0116] Optionally, the initial node attribute information detection model also includes: a first maximization processing module, a second maximization processing module, and a minimization processing module.

[0117] The processing unit 603 is further configured to input the topology view data and topology feature map into the first maximization processing module to obtain the corresponding topology feature information; input the attribute view data and attribute feature map into the second maximization processing module to obtain the corresponding attribute feature information; input the topology feature map and attribute feature map into the minimization processing module to obtain fused semantic information; input the topology feature information, attribute feature information, and fused semantic information into the feature fusion network to obtain fused feature information; and input the fused feature information into the linear classification network model to obtain the predicted attribute information corresponding to each sample node.

[0118] Figure 7 This is a schematic diagram of the structure of a node attribute information detection device 70 in a trading platform provided by an embodiment of the present invention. The detection device may include:

[0119] Acquisition unit 701 is used to acquire graph network data of the trading platform to be tested.

[0120] The processing unit 702 is used to input the graph network data of the trading platform to be detected into the node attribute information detection model to obtain the attribute information corresponding to each node in the trading platform to be detected.

[0121] The attribute information includes whether the node is abnormal or normal.

[0122] Optionally, the node attribute information detection model includes: decoupled network, first multilayer perceptron network, second multilayer perceptron network, feature fusion network and linear classification network.

[0123] Processing unit 702 is specifically used to input the graph network data of the transaction platform to be detected into a decoupled network to obtain the topology view data and attribute view data corresponding to the transaction platform to be detected; input the topology view data and attribute view data corresponding to the transaction platform to be detected into a first multilayer perceptron network and a second multilayer perceptron network, respectively. The first multilayer perceptron network and the second multilayer perceptron network are used to perform dimensionality reduction processing on the topology view data and attribute view data, respectively, to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data; input the topology feature map and attribute feature map into a feature fusion network to obtain the fused feature information corresponding to the transaction platform to be detected; finally, input the fused feature information corresponding to the transaction platform to be detected into a linear classification network model to obtain the attribute information corresponding to each node in the transaction platform to be detected.

[0124] Optionally, the node attribute information detection model may also include: a first maximization processing module, a second maximization processing module, and a minimization processing module.

[0125] The processing unit 702 is further configured to input the topology view data and topology feature map corresponding to the transaction platform to be detected into the first maximization processing module to obtain the corresponding topology feature information; input the attribute view data and attribute feature map into the second maximization processing module to obtain the corresponding attribute feature information; input the topology feature map and attribute feature map into the minimization processing module to obtain fused semantic information; and input the topology feature information, attribute feature information, and fused semantic information into the feature fusion network to obtain the fused feature information corresponding to the transaction platform to be detected.

[0126] Figure 8 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as... Figure 8As shown, the electronic device may include a processor 810, a communication interface 820, a memory 830, and a communication bus 840. The processor 810, communication interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions from the memory 830 to execute the training method for the node attribute information detection model in the trading platform, or the node attribute information detection method in the trading platform.

[0127] The training method for the node attribute information detection model in the trading platform may include: acquiring graph network data of the trading platform; selecting some nodes in the graph network data as sample nodes and labeling the true attribute information of each sample node; the attribute information of the nodes includes abnormal nodes or normal nodes; inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node; updating the model parameters of the initial node attribute information detection model according to the true attribute information and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model.

[0128] A method for detecting node attribute information in a trading platform may include: acquiring graph network data of the trading platform to be tested; inputting the graph network data of the trading platform to be tested into a node attribute information detection model to obtain the attribute information corresponding to each node in the trading platform to be tested. The node attribute information detection model is the node attribute information detection model described in the above embodiments.

[0129] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0130] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the training method for the node attribute information detection model in the trading platform provided by the above methods, or the node attribute information detection method in the trading platform.

[0131] The training method for the node attribute information detection model in the trading platform may include: acquiring graph network data of the trading platform; selecting some nodes in the graph network data as sample nodes and labeling the true attribute information of each sample node; the attribute information of the nodes includes abnormal nodes or normal nodes; inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node; updating the model parameters of the initial node attribute information detection model according to the true attribute information and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model.

[0132] The method for detecting node attribute information in a trading platform may include: obtaining graph network data of the trading platform to be detected; inputting the graph network data of the trading platform to be detected into the node attribute information detection model to obtain the attribute information corresponding to each node in the trading platform to be detected.

[0133] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a training method for a node attribute information detection model in a trading platform, or a node attribute information detection method in a trading platform, provided by the methods described above.

[0134] The training method for the node attribute information detection model in the trading platform may include: acquiring graph network data of the trading platform; selecting some nodes in the graph network data as sample nodes and labeling the true attribute information of each sample node; the attribute information of the nodes includes abnormal nodes or normal nodes; inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node; updating the model parameters of the initial node attribute information detection model according to the true attribute information and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model.

[0135] The method for detecting node attribute information in a trading platform may include: obtaining graph network data of the trading platform to be detected; inputting the graph network data of the trading platform to be detected into the node attribute information detection model to obtain the attribute information corresponding to each node in the trading platform to be detected.

[0136] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0137] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training method for a node attribute information detection model in a trading platform, characterized in that, include: Obtain graph network data from the trading platform; A subset of nodes in the graph network data are selected as sample nodes, and the true attribute information of each sample node is labeled; the attribute information of the node includes abnormal nodes or normal nodes; The graph network data is input into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node; The model parameters of the initial node attribute information detection model are updated based on the actual attribute information and predicted attribute information of the sample nodes to obtain the trained node attribute information detection model. The initial node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network; The step of inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node includes: The graph network data is input into the decoupled network to obtain the topology view data and attribute view data corresponding to the graph network data; The topology view data is input into the first multilayer perceptron network, and the attribute view data is input into the second multilayer perceptron network. The first multilayer perceptron network and the second multilayer perceptron network are respectively used to perform dimensionality reduction processing on the topology view data and the attribute view data to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data. The topological feature map and attribute feature map are input into the feature fusion network to obtain the fused feature information; The fused feature information is input into the linear classification network model to obtain the predicted attribute information corresponding to each sample node.

2. The training method for the node attribute information detection model in the trading platform according to claim 1, characterized in that, The initial node attribute information detection model further includes: a first maximization processing module, a second maximization processing module, and a minimization processing module; The step of inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node further includes: The topology view data and topology feature map are input into the first maximization processing module to obtain the corresponding topology feature information; the attribute view data and attribute feature map are input into the second maximization processing module to obtain the corresponding attribute feature information. The topological feature map and attribute feature map are input into the minimization processing module to obtain fused semantic information; The topological feature information, attribute feature information, and fused semantic information are input into the feature fusion network to obtain the fused feature information.

3. The training method for the node attribute information detection model in the trading platform according to claim 1 or 2, characterized in that, The loss function of the initial node attribute information detection model is as follows (1): in, In formula (2), v i This represents the i-th sample node. Let y represent the set of sample nodes. i Let W represent the true label of the i-th node, σ represent the non-linear activation function, and W represent the true label of the i-th node. ′ This represents the weight matrix of a linear classification network. b represents the learned features of the i-th node. ′ This represents the bias vector of the linear classification network; In formulas (3) and (4), x1 represents the index of the topological feature information, x2 represents the index of the attribute feature information, z1 represents the topological feature map, and z2 represents the attribute feature map; q(x i |z i ) represents x i and z i Variational estimation under conditional distributions, where p(z1,x2) represents the joint probability distribution of z1 and x2, E represents the expectation, r(z1) represents the marginal distribution of z1, and r(z2) represents the marginal distribution of z2. Let z2 represent the probability distribution of z2 under the condition x2 in the feature space of x2; This indicates that in the feature space of x1, z1 lies in x. 21 The probability distribution under the given conditions.

4. A method for detecting node attribute information in a trading platform, characterized in that, include: Obtain the graph network data of the trading platform to be tested; The graph network data of the transaction platform to be tested is input into the node attribute information detection model to obtain the attribute information corresponding to each node in the transaction platform to be tested; the attribute information includes abnormal nodes or normal nodes. The node attribute information detection model is the node attribute information detection model described in any one of claims 1-3.

5. The method for detecting node attribute information in a trading platform according to claim 4, characterized in that, The node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network; The step of inputting the graph network data of the transaction platform to be tested into the node attribute information detection model to obtain the attribute information corresponding to each node in the transaction platform to be tested includes: The graph network data of the transaction platform to be tested is input into the decoupled network to obtain the topology view data and attribute view data corresponding to the transaction platform to be tested. The topology view data and attribute view data corresponding to the transaction platform to be detected are respectively input into the first multilayer perceptron network and the second multilayer perceptron network; The first multilayer perceptron network and the second multilayer perceptron network are respectively used to perform dimensionality reduction processing on the topology view data and the attribute view data to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data. The topological feature map and attribute feature map are input into the feature fusion network to obtain the fused feature information corresponding to the transaction platform to be detected. The fused feature information corresponding to the transaction platform to be detected is input into the linear classification network model to obtain the attribute information corresponding to each node in the transaction platform to be detected.

6. The method for detecting node attribute information in a trading platform according to claim 5, characterized in that, The node attribute information detection model further includes: a first maximization processing module, a second maximization processing module, and a minimization processing module; The step of inputting the graph network data of the transaction platform to be tested into the node attribute information detection model to obtain the attribute information corresponding to each node in the transaction platform to be tested also includes: The topology view data and topology feature map corresponding to the transaction platform to be detected are input into the first maximization processing module to obtain the corresponding topology feature information; the attribute view data and attribute feature map are input into the second maximization processing module to obtain the corresponding attribute feature information. The topological feature map and attribute feature map are input into the minimization processing module to obtain fused semantic information; The topological feature information, attribute feature information, and fused semantic information are input into the feature fusion network to obtain the fused feature information corresponding to the transaction platform to be detected.

7. The method for detecting node attribute information in a trading platform according to any one of claims 4-6, characterized in that, The detection method further includes: Construct the graph network corresponding to the transaction platform to be tested; Feature extraction is performed on the graph network corresponding to the transaction platform to be tested to obtain the graph network data of the transaction platform to be tested.

8. A training device for a node attribute information detection model in a trading platform, characterized in that, include: The acquisition unit is used to acquire graph network data from the trading platform; The sample annotation unit is used to select a portion of the nodes in the graph network data as sample nodes and to annotate the true attribute information of each sample node; the attribute information of the node includes abnormal nodes or normal nodes. The processing unit is used to input the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node; The update unit is used to update the model parameters of the initial node attribute information detection model according to the real attribute information and predicted attribute information of the sample node, so as to obtain the trained node attribute information detection model. The initial node attribute information detection model includes: a decoupled network, a first multilayer perceptron network, a second multilayer perceptron network, a feature fusion network, and a linear classification network; The step of inputting the graph network data into the initial node attribute information detection model to obtain the predicted attribute information corresponding to each sample node includes: The graph network data is input into the decoupled network to obtain the topology view data and attribute view data corresponding to the graph network data; The topology view data is input into the first multilayer perceptron network, and the attribute view data is input into the second multilayer perceptron network. The first multilayer perceptron network and the second multilayer perceptron network are respectively used to perform dimensionality reduction processing on the topology view data and the attribute view data to obtain the topology feature map corresponding to the topology view data and the attribute feature map corresponding to the attribute view data. The topological feature map and attribute feature map are input into the feature fusion network to obtain the fused feature information; The fused feature information is input into the linear classification network model to obtain the predicted attribute information corresponding to each sample node.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method for the node attribute information detection model in the trading platform as described in any one of claims 1 to 3; or, it implements the node attribute information detection method in the trading platform as described in any one of claims 4 to 7.

Images