Block chain address behavior classification method and system based on graph neural network

Through the blockchain address behavior classification method based on graph neural network, the problem of inefficiency in cryptocurrency supervision is solved, efficient and accurate address behavior classification is achieved, and regulatory efficiency and data analysis capabilities are improved.

CN118965060APending Publication Date: 2024-11-15ZHEJIANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410965090.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-15

Smart Images

  • Figure CN118965060A_ABST
    Figure CN118965060A_ABST
Patent Text Reader

Abstract

The invention discloses a block chain address behavior classification method and system based on a graph neural network, and belongs to the field of block chain data analysis. The method specifically comprises the following steps: constructing an address transaction graph list corresponding to each block chain address by utilizing block chain address transaction data; the address transaction graph list comprises a plurality of address transaction graphs, learning features of each address transaction graph by using a graph neural network, and generating graph embedding features of each address transaction graph; and aggregating all the graph embedding features of one block chain address, and obtaining a prediction result of block chain address behavior classification by using the classification model. According to the method, a novel block chain address behavior classification scheme is provided, and the addresses in the block chain network can be accurately classified.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of blockchain data analysis, and specifically relates to a blockchain address behavior classification method and system based on graph neural network. Background Art

[0002] The concept of blockchain was first proposed by Satoshi Nakamoto in 2008. In his white paper, he introduced in detail the underlying technology of blockchain, a cryptocurrency. Cryptocurrency uses blockchain, a distributed ledger technology, to store and transfer assets in a peer-to-peer network without relying on the participation of a trusted third party. Blockchain is a decentralized trust network based on cryptography and distributed computing. Through the joint participation and verification of multiple nodes, it achieves reliable recording and protection of transactions. This innovative technical architecture gives blockchain unique characteristics, such as decentralization, immutability and transparency, which makes it have broad application prospects in finance, supply chain management, the Internet of Things and other fields.

[0003] Although blockchain technology has many potential advantages, it also faces a series of challenges. Among them, the decentralized and anonymous characteristics of cryptocurrencies have brought great obstacles to regulatory work. Cryptocurrency transactions are usually conducted anonymously or pseudo-anonymously, which makes it difficult for legal and regulatory authorities to combat money laundering, terrorist financing, and other illegal activities. The anonymity of transactions complicates the tracking and accountability of illegal activities. In order to effectively deal with the illegal activities brought about by cryptocurrency technology, more innovative regulatory technologies need to be adopted and appropriate legal and regulatory frameworks need to be formulated. This includes the introduction of new regulatory tools and technologies, such as cryptocurrency user behavior analysis and identity recognition technology, to enhance the traceability and transparency of cryptocurrency transactions by regulators. These new technologies can help regulators identify and analyze cryptocurrency transactions involving illegal activities and accurately match the identities of the parties involved, thereby improving regulatory effectiveness.

[0004] At present, the existing cryptocurrency user behavior analysis and identity recognition technologies face some significant shortcomings. First, the existing technical means are relatively backward and rely more on rules rather than advanced artificial intelligence algorithms such as deep learning. This rule-based analysis method is powerless when dealing with complex and changeable trading behaviors. Secondly, the existing technology is difficult to adapt to large-scale cryptocurrency address analysis. With the rapid growth of cryptocurrency trading volume and the number of users, the existing technology often encounters performance bottlenecks when processing and analyzing massive data, and it is difficult to provide timely and accurate analysis results. Finally, the existing technology also has shortcomings in the automatic identification and classification of cryptocurrency addresses. The current system relies more on human intervention and manual operation, which is inefficient and error-prone. These shortcomings limit the application effect of existing technologies in large-scale and highly complex cryptocurrency supervision.

[0005] The number of active cryptocurrency addresses has increased tenfold in the past decade. The new generation of regulatory technology needs to combine deep learning and artificial intelligence algorithms to improve the ability to identify complex transaction behaviors. At the same time, it is also necessary to develop efficient data processing and analysis technologies to meet the needs of large-scale data analysis. In addition, the introduction of automated identification and classification systems can significantly improve efficiency and reduce human errors. In order to strengthen the monitoring and tracking of illegal cryptocurrency transactions and improve regulatory effectiveness, an accurate and efficient address classification method is needed. Summary of the invention

[0006] In order to solve the problems in the prior art, the present invention provides a blockchain address behavior classification method and system based on graph neural network.

[0007] The technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention discloses a blockchain address behavior classification method based on a graph neural network, comprising the following steps:

[0009] 1) Use blockchain address transaction data to build an address transaction graph list corresponding to each blockchain address;

[0010] 2) The address transaction graph list includes multiple address transaction graphs. The graph neural network is used to learn the features of each address transaction graph and generate graph embedding features of each address transaction graph.

[0011] 3) Aggregate all graph embedding features of the blockchain address to be classified and use the classification model to obtain the prediction result of the behavior classification of the blockchain address.

[0012] Furthermore, the step 1) is specifically as follows:

[0013] Divide the blockchain address transaction data into multiple groups in the order of timestamps. The blockchain address transaction data in each group are equal. Process the blockchain address transaction data in each group. All blockchain addresses involved in the blockchain address transaction data are used as address nodes. Address nodes are connected to each other through corresponding transactions. Construct an initial address transaction graph corresponding to each blockchain address.

[0014] First, compress and merge the address nodes that only involve one transaction to obtain a compressed address transaction graph, and then compress and merge the address nodes that involve multiple transactions in the compressed address transaction graph to obtain a compressed address transaction graph;

[0015] Perform centrality calculation on the address nodes in the compressed address transaction graph to obtain the centrality features of the address nodes, and integrate the centrality features into the corresponding address nodes to obtain the final address transaction graph;

[0016] All final address transaction graphs of a blockchain address constitute the address transaction graph list corresponding to the blockchain address.

[0017] Furthermore, after the address nodes involving only one transaction are compressed and merged, the address features of the merged address nodes are calculated; the address features include data sample information, dispersion degree measurement, central tendency measurement and data distribution form measurement;

[0018] After compressing and merging the address nodes involved in multiple transactions in a compressed address transaction graph, the address features of the merged address nodes will also be calculated.

[0019] Furthermore, the step 2) is specifically as follows:

[0020] A graph convolution is performed on an address transaction graph through a graph neural network to obtain the feature representation of the address transaction graph. The feature representation of the address transaction graph is then processed using a multi-layer perceptron to obtain the node representation of the address transaction graph. The node representation of the address transaction graph is aggregated through a readout function to obtain the overall representation of the address transaction graph. All node embeddings in the overall representation of the address transaction graph are then aggregated to obtain graph embedding features.

[0021] Furthermore, the step 3) is specifically as follows:

[0022] All graph embedding features of the blockchain address to be classified are aggregated using a long short-term memory network, and the aggregation results are processed using a multi-layer perceptron to obtain the prediction result of the blockchain address behavior classification.

[0023] In a second aspect, the present invention discloses a blockchain address behavior classification system for implementing the method, comprising:

[0024] An address transaction graph list construction module, which is used to construct an address transaction graph list corresponding to each blockchain address using blockchain address transaction data;

[0025] A graph embedding feature acquisition module is used to generate graph embedding features of each address transaction graph by using a graph neural network to learn the features of each address transaction graph;

[0026] The address classification module is used to aggregate all graph embedding features of the blockchain address to be classified and use the classification model to obtain the prediction results of the blockchain address behavior classification.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] (1) This paper proposes for the first time a blockchain address behavior classification method based on graph neural network, and proposes the first accurate and efficient solution to this problem. This paper innovatively uses graph neural network to extract features from the blockchain transaction graph structure.

[0029] (2) This paper proposes a novel blockchain address behavior classification system that can automatically extract blockchain address behavior features and classify addresses. The system is designed with three key components, namely address graph construction, graph representation learning, and address behavior classification.

[0030] (3) The present invention innovatively introduces graph node compression technology and graph structure enhancement technology to convert transactions of blockchain addresses into a unified graph representation for easy input into the graph neural network for processing.

[0031] (4) The present invention innovatively adopts graph neural network to extract address graph features, and selects the combination of LSTM (Long Short Term Memory) + MLP (Multilayer Perceptron) as the behavior classification model, thereby realizing address classification.

[0032] (5) The framework and method proposed in this invention have been verified in a dataset of more than one million. Through experimental research, the method proposed in this invention has a significant effect in identifying blockchain address behavior, with an accuracy and F1 score of 96% and 95% respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is the overall design structure diagram of the present invention.

[0034] Figure 2 Infographic of the statistical feature extractor of the present invention.

[0035] Figure 3 Schematic diagram of the single transaction address compression process of the present invention.

[0036] Figure 4 Schematic diagram of the multi-transaction address compression process of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described and illustrated below in conjunction with specific embodiments. The embodiments are merely exemplary of the present disclosure and do not define the scope of limitation. The technical features of each embodiment of the present invention may be combined accordingly without conflicting with each other.

[0038] The overall design structure diagram of the present invention is as follows Figure 1As shown in the figure, the focus is divided into three parts: address transaction graph construction, address transaction graph representation learning, and address classification. The address transaction graph construction part includes original graph construction, graph node compression, and graph structure enhancement; the address transaction graph representation learning part includes graph feature convolution, node representation learning, and graph feature extraction; the address classification part uses the LSTM model.

[0039] The first step: address transaction graph construction

[0040] Component 1: Original graph construction. Since transactions in blockchain usually exist in a serialized form, the present invention implements a component that converts serialized transactions into an address transaction graph list.

[0041] Before building the address graph, the system needs to group and slice the transactions of the address. This is because the number of transactions of different addresses varies greatly, and building all transactions of all addresses in one graph will result in very different scales of the graph. At the same time, building all transactions in one graph cannot reflect the timing characteristics between transactions, resulting in feature loss. To solve this problem, the system divides the transactions of the address into multiple groups in the order of timestamps, with a fixed number of transactions in each group. The system ultimately chooses to contain 100 transactions per slice unit, which means that each graph contains information on 100 transactions.

[0042] After completing the grouping of transactions, the system will process each group of transactions and construct the corresponding address transaction graph. The system extracts all addresses involved in each group of transactions as points, and connects the points through corresponding transactions to obtain the corresponding address transaction graph. Ultimately, all transactions of an address can be described as a list consisting of a series of address transaction graphs.

[0043] Component 2: Graph node compression. Through the original graph construction module, the blockchain can convert the transactions of blockchain addresses into address graph lists. However, due to the size of the graph, the constructed original address graph cannot be directly input into the graph neural network for processing, and graph node compression is required.

[0044] Graph node compression includes two compression processes: one is single transaction address node, this type of address only involves one transaction, that is, the node degree is 1; the other is multi-transaction address node, this type of address involves multiple transactions, that is, the node degree is greater than 1. In order to compress address nodes, the system introduces a two-stage compression algorithm, including single transaction address compression and multi-transaction address compression.

[0045] In order to retain the statistical characteristics of transactions while compressing nodes, this system designs a statistical feature extraction method to capture the transfer characteristics of each transaction. Figure 2The statistical features included in this method are shown. The statistical feature extraction method extracts four types of features, namely: data sample information, dispersion measurement, central tendency measurement, and data distribution morphology measurement. Since most of the available information in blockchain transactions is the transfer amount, when merging nodes, this method will extract and process the transfer amount of the nodes to be merged, and use these features as the features of the merged address.

[0046] Single transaction address compression aims to merge address nodes with only a single transaction, thereby reducing the number of address nodes in the graph. Figure 3 The specific process is shown. For single transaction address nodes connected to the same transaction node, the system merges these nodes into a super node, called a single transaction super node.

[0047] In order to retain the input or output amounts of these addresses, the system uses statistical feature extraction methods to extract the transfer amount features of these nodes that need to be merged, and generates and saves them as the features of a single transaction super node.

[0048] After completing the single transaction address compression, a transaction node can be connected to at most two single transaction super nodes as input and output respectively.

[0049] Different from the single transaction address compression algorithm that can directly merge the address nodes that are only connected to the transaction node by one edge, the multi-transaction address compression needs to first compare the similarity of all multi-transaction address nodes to determine the merged object. In order to achieve similarity comparison, the system constructs the node adjacency matrix, and each element in the matrix represents an edge in the graph; first, using its own inverse matrix and normalizing the matrix itself, the system can obtain the similarity matrix of the matrix; then, after obtaining the similarity matrix, the system sets a screening threshold (0.7 in this system) to filter the elements in the matrix, and only retains the matrix elements greater than the screening threshold; finally, the non-zero element nodes in each row of the matrix will be merged, and the merging method will adopt multi-transaction address compression. Figure 4 The multi-transaction address compression process is demonstrated. When the similarity of multiple multi-transaction address nodes exceeds the screening threshold set by the system, these address nodes are merged into a multi-transaction super node, and the transfer amount features of the multi-transaction super node are calculated using the same statistical feature extraction method as in single transaction address compression.

[0050] In a specific embodiment of the present invention, the graph node compression is as follows: first, the address nodes involving only one transaction are compressed and merged, wherein the address nodes involving only one transaction include single transaction input address nodes and single transaction output address nodes, and the address nodes involving only one transaction are compressed and merged as follows: single transaction input address nodes for the same transaction are merged and single transaction output address nodes for the same transaction are merged to obtain a compressed address transaction graph;

[0051] The address nodes involved in multiple transactions in a compressed address transaction graph are compressed and merged, wherein the address nodes involved in multiple transactions include multiple transaction input address nodes and multiple transaction output address nodes, similarity calculation is performed on all multiple transaction input address nodes, and similarity calculation is performed on all multiple transaction output address nodes. When the similarity of multiple multiple transaction input address nodes or the similarity of multiple multiple transaction output address nodes exceeds a set threshold, these multiple transaction input address nodes or multiple transaction output address nodes are merged to obtain a compressed address transaction graph.

[0052] Among them, similarity calculation is performed on all multi-transaction input address nodes and similarity calculation is performed on all multi-transaction output address nodes. When the similarity of multiple multi-transaction input address nodes or the similarity of multiple multi-transaction output address nodes exceeds the set threshold, these multi-transaction input address nodes or multi-transaction output address nodes are merged to obtain a compressed address transaction graph. Specifically, the adjacency matrix of the multi-transaction input address nodes or multi-transaction output address nodes is first constructed, and then the adjacency matrix is ​​normalized using the inverse matrix of the adjacency matrix to obtain the similarity matrix of the adjacency matrix; the elements in the similarity matrix are filtered by a preset screening threshold, and only the matrix elements greater than the screening threshold are retained, and the matrix elements less than the screening threshold are set to 0, and the address nodes corresponding to the non-zero elements in each row of the filtered similarity matrix are merged, and finally, the compressed address transaction graph is obtained.

[0053] Component 3: Graph structure enhancement. Through the graph node compression process, the system can obtain a compressed address graph set. In order to further enhance the node structure information of each graph in the set, the system proposes a graph structure enhancement method based on network centrality to enhance the structural information of address nodes in the graph. The system calculates four types of network centrality of the graph, namely degree centrality, betweenness centrality, closeness centrality, and PageRank centrality. After the system completes the calculation of the centrality of the graph, it integrates the calculated features into the graph nodes, that is, splicing the calculated features to the original features. Through graph structure enhancement, the system can further expand the node features of the address graph so that it contains both the feature information of the address transaction amount data and the enhanced graph structure features covering the graph network centrality.

[0054] Second step: address transaction graph representation learning

[0055] The present invention uses graph neural network (GFN) to realize address transaction graph representation learning. GFN can learn high-order node representation, which is very suitable for processing node features and graph structure in address transaction graph. The address transaction graph representation learning module includes three parts, namely graph feature convolution, node representation learning and graph feature extraction.

[0056] Component 1: Graph Feature Convolution. Graph neural networks extract feature representations of graphs through the graph convolution process. In each graph convolution layer, the features of a node and its neighboring nodes are aggregated to obtain first-order node embeddings, thereby achieving higher-order node embeddings. The system uses the following two types of information to represent nodes: one is the structural features of the graph, which mainly include the degree and centrality of the node; the other is the propagation features of the graph, which are mainly extracted from the high-order adjacency matrix. After these input features are convolved with multi-scale graphs, high-order node embedding representations are generated and sent to node representation learning for processing.

[0057] Component 2: Node representation learning. After graph feature convolution, the system obtains the feature representation of the graph. Afterwards, the system uses a multi-layer perceptron (MLP) to capture more effective graph features. By connecting the MLP layer in series with the label of each graph, the system can train a classification model to provide a more abstract representation of the nodes in the graph. The MLP layer further processes the input features through nonlinear activation functions and trainable parameters to capture higher-level node representations. Unlike traditional graph convolutional networks, GFN removes the adjacency matrix and uses neural networks to obtain higher-level node representations, which improves computational efficiency while retaining more feature information.

[0058] Component 3: Graph feature extraction. Finally, the system extracts graph features through the readout function. The readout function generates the overall representation of the graph by aggregating the representations of all nodes in the graph. The system uses the global aggregation function SUM to aggregate the embeddings of all nodes in the graph to obtain the final graph embedding features. These graph embedding features can be used for subsequent classification or other tasks, providing a global representation of the graph.

[0059] The third step: address classification

[0060] So far, this system has realized the mapping of address graph to graph embedding with the help of graph neural network. Now, the system considers representing the transaction features of a specific address as an embedding list with time features. Since the number of transactions on each address is different, the number of elements in the embedding list corresponding to each address is also different, but the transaction behavior in the address is fixed, which corresponds to a clear address label. Therefore, the goal of the address classification stage is to determine the label of the address by analyzing the embedding list of the address, which can be transformed into the problem of classifying a vector list of uncertain length. To achieve this goal, this system uses a long short-term memory network (LSTM) to process the embedding list.

[0061] Long Short-Term Memory (LSTM) is a special type of recurrent neural network (RNN) that is suitable for processing and predicting long-term dependencies in time series data. By introducing memory units and gating mechanisms, LSTM can effectively capture and retain dependencies over long time spans, thereby solving the gradient vanishing and gradient exploding problems of traditional RNNs when processing long sequences. Therefore, LSTM has a wide range of applications in fields such as processing time series data and natural language processing.

[0062] By using LSTM, the system can convert the list of graphical representations of addresses into a vector. To obtain the final classification result, the system connects a multi-layer perceptron (MLP) layer after the LSTM layer to classify the list of graphical representations. By training the model, the system can optimize the relevant parameters of LSTM and MLP to achieve accurate address classification. During the training process, the system uses supervised learning methods to continuously adjust the model parameters through labeled data sets to improve the accuracy and robustness of classification.

[0063] In the embodiment of the present invention, the blockchain address transaction data set used is a cryptocurrency address transaction data set, which contains more than 2 million cryptocurrency addresses and their transactions. The present invention uses an automatic crawling tool to create the data set by collecting cryptocurrency addresses from different sources, including websites such as btc.com, walletexplorer.com, and oxt.me. In addition, the collected data is manually double-checked to ensure data quality.

[0064] In order to evaluate the performance of the system, the present invention selects accuracy and F1 score as evaluation indicators. Accuracy refers to the proportion of samples correctly classified by the classifier to the total number of samples, while the F1 score is the harmonic mean of precision and recall, which comprehensively reflects the performance of the classifier. In the system verification, the present invention uses the constructed data set for testing, and the final accuracy and F1 score reach 96% and 95% respectively. This evaluation result shows that the present invention has shown high effectiveness and reliability in the classification of cryptocurrency addresses.

[0065] The above-mentioned embodiments only express several implementation methods of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. For ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A blockchain address behavior classification method based on graph neural network, It is characterized in that The steps include: 1) Use blockchain address transaction data to build an address transaction graph list corresponding to each blockchain address; 2) The address transaction graph list includes multiple address transaction graphs. The graph neural network is used to learn the features of each address transaction graph and generate graph embedding features of each address transaction graph. 3) Aggregate all graph embedding features of the blockchain address to be classified and use the classification model to obtain the prediction result of the behavior classification of the blockchain address.

2. The method according to claim 1, It is characterized in that The step 1) is specifically as follows: Divide the blockchain address transaction data into multiple groups in the order of timestamps. The blockchain address transaction data in each group are equal. Process the blockchain address transaction data in each group. All blockchain addresses involved in the blockchain address transaction data are used as address nodes. Address nodes are connected to each other through corresponding transactions. Construct an initial address transaction graph corresponding to each blockchain address. First, compress and merge the address nodes that only involve one transaction to obtain a compressed address transaction graph, and then compress and merge the address nodes that involve multiple transactions in the compressed address transaction graph to obtain a compressed address transaction graph; Perform centrality calculation on the address nodes in the compressed address transaction graph to obtain the centrality features of the address nodes, and integrate the centrality features into the corresponding address nodes to obtain the final address transaction graph; All final address transaction graphs of a blockchain address constitute the address transaction graph list corresponding to the blockchain address.

3. The method according to claim 2, It is characterized in that The address nodes involving only one transaction are first compressed and merged to obtain a compressed address transaction graph, and then the address nodes involving multiple transactions in the compressed address transaction graph are compressed and merged to obtain a compressed address transaction graph; specifically: First, the address nodes involving only one transaction are compressed and merged, wherein the address nodes involving only one transaction include single transaction input address nodes and single transaction output address nodes, and the single transaction input address nodes for the same transaction are merged and the single transaction output address nodes for the same transaction are merged to obtain a compressed address transaction graph; The address nodes involved in multiple transactions in a compressed address transaction graph are compressed and merged, wherein the address nodes involved in multiple transactions include multiple transaction input address nodes and multiple transaction output address nodes, similarity calculation is performed on all multiple transaction input address nodes, and similarity calculation is performed on all multiple transaction output address nodes. When the similarity of multiple multiple transaction input address nodes or the similarity of multiple multiple transaction output address nodes exceeds a set threshold, these multiple transaction input address nodes or multiple transaction output address nodes are merged to obtain a compressed address transaction graph.

4. The method according to claim 3, It is characterized in that After the address nodes are compressed and merged, the address features of the merged address nodes are calculated; the address features include data sample information, dispersion degree measurement, central tendency measurement and data distribution shape measurement.

5. The method according to claim 3, It is characterized in that The similarity calculation is performed on all multi-transaction input address nodes and all multi-transaction output address nodes. When the similarity of multiple multi-transaction input address nodes or the similarity of multiple multi-transaction output address nodes exceeds the set threshold, these multi-transaction input address nodes or multi-transaction output address nodes are merged to obtain a compressed address transaction graph; specifically: First, an adjacency matrix of multiple transaction input address nodes or multiple transaction output address nodes is constructed, and then the adjacency matrix is ​​normalized using the inverse matrix of the adjacency matrix to obtain a similarity matrix of the adjacency matrix; The elements in the similarity matrix are filtered by a preset filtering threshold, and only the matrix elements greater than the filtering threshold are retained. The matrix elements less than the filtering threshold are set to 0, and the address nodes corresponding to the non-zero elements in each row of the filtered similarity matrix are merged. Finally, a compressed address transaction graph is obtained.

6. The method according to claim 2, It is characterized in that The centrality calculation includes degree centrality calculation, betweenness centrality calculation, closeness centrality calculation and PageRank centrality calculation.

7. The method according to claim 1, It is characterized in that The step 2) is specifically as follows: A graph convolution is performed on an address transaction graph through a graph neural network to obtain the feature representation of the address transaction graph. The feature representation of the address transaction graph is then processed using a multi-layer perceptron to obtain the node representation of the address transaction graph. The node representation of the address transaction graph is aggregated through a readout function to obtain the overall representation of the address transaction graph. All node embeddings in the overall representation of the address transaction graph are then aggregated to obtain graph embedding features.

8. The method according to claim 1, It is characterized in that The step 3) is specifically as follows: All graph embedding features of the blockchain address to be classified are aggregated using a long short-term memory network, and the aggregation results are processed using a multi-layer perceptron to obtain the prediction result of the blockchain address behavior classification.

9. A blockchain address behavior classification system implementing the method of claim 1, It is characterized in that include: An address transaction graph list construction module, which is used to construct an address transaction graph list corresponding to each blockchain address using blockchain address transaction data; A graph embedding feature acquisition module is used to generate graph embedding features of each address transaction graph by using a graph neural network to learn the features of each address transaction graph; The address classification module is used to aggregate all graph embedding features of the blockchain address to be classified and use the classification model to obtain the prediction results of the blockchain address behavior classification.