Data violation resale detection method and system based on enhanced relational graph convolutional network

Abstract

This invention provides a method and system for detecting illegal data resale based on an enhanced relational graph convolutional network. The method includes: acquiring data transaction information from multiple platforms; constructing a multi-relational transaction topology graph with each transaction as a node and including semantic edges of role sharing and role exchange; inputting the multi-relational transaction topology graph into a data illegal resale detection model based on an enhanced relational graph convolutional network to detect illegal resale in the transaction information; and outputting the illegal resale results from the transaction information using the enhanced relational graph convolutional network-based data illegal resale detection model. This invention can accurately adapt to the characteristics of multi-relational transaction topology, effectively improving the accuracy and efficiency of data illegal resale detection.

Description

Technical Field

[0001] This invention relates to the field of data security detection and graph neural network application technology, specifically to a multi-platform data illegal resale detection method and system based on enhanced relational graph convolutional networks. Background Technology

[0002] With the rapid development of the digital economy, data transactions on multiple platforms are becoming increasingly frequent, and the value of data as a major factor of production is constantly increasing. However, the illegal resale of data is also becoming more prominent, seriously infringing on the rights and interests of data subjects, disrupting the order of the data trading market, and even triggering data security risks.

[0003] Currently, existing methods for detecting illegal resale of data mostly rely on traditional rule matching or simple machine learning models, which are difficult to handle the complex relationships between entities in multi-platform transactions. Some detection methods based on graph neural networks fail to fully explore the differences in the roles of the transaction entities, cannot adapt to multi-relationship transaction scenarios, and have limited feature extraction capabilities, making it difficult to accurately capture the semantic context and temporal causal relationships of transactions. This results in low detection accuracy and high false negative and false positive rates, failing to meet the actual needs of illegal data transaction detection in multi-platform transactions. Summary of the Invention

[0004] This invention provides a data resale violation detection method based on an enhanced relational graph convolutional network, comprising: acquiring data transaction information from multiple platforms; constructing a multi-relational transaction topology graph with each transaction as a node and including semantic edges of role sharing and role exchange; and inputting the multi-relational transaction topology graph into a data resale violation detection model based on an enhanced relational graph convolutional network to detect the existence of resale violations in the transaction information.

[0005] According to one embodiment of the present invention, the acquisition of multi-platform data transaction information specifically includes: parsing the acquired multi-platform data transaction information, extracting the transaction index, buyer identity identifier, and seller identity identifier corresponding to each transaction; separating the buyer identity identifier set, seller identity identifier set, and mixed identity identifier set from the parsed transaction information; the mixed identity identifier set is the union of the buyer identity identifier set and the seller identity identifier set; defining an edge creation function, the input of which is the identity identifier set and the parsed transaction information, and the output is an undirected edge set; the execution logic of the edge creation function is as follows: traversing each identity identifier in the input identity identifier set, matching all transaction indices in which the identity identifier participates as a corresponding role in the transaction information, performing pairwise undirected connections on all transaction indices associated with the same identity identifier to generate undirected edges, and summarizing them into an edge set.

[0006] According to one embodiment of the present invention, constructing a transaction topology graph with role-shared semantic edges includes: initializing an empty set of role-shared edges; inputting the set of buyer identity identifiers into the edge creation function to obtain a subset of undirected edges shared by buyer roles, and adding the subset to the set of role-shared edges; inputting the set of seller identity identifiers into the edge creation function to obtain a subset of undirected edges shared by seller roles, and adding the subset to the set of role-shared edges, thereby completing the construction of the set of role-shared edges; all undirected edges in the set of role-shared edges are role-shared semantic edges in the multi-relationship transaction topology graph.

[0007] According to one embodiment of the present invention, constructing a transaction topology graph with role-swapping semantic edges includes: initializing an empty set of mixed undirected edges and an empty set of role-swapping edges; inputting the set of mixed identity identifiers into the edge creation function to obtain a set of mixed undirected edges associated with all transaction entities; performing a set operation on the set of mixed undirected edges to remove all undirected edges contained in the set of role-shared edges, and summing the remaining undirected edges to form a set of role-swapping edges; all undirected edges in the set of role-swapping edges are role-swapping semantic edges in the multi-relationship transaction topology graph.

[0008] According to one embodiment of the present invention, the method further includes constructing the data violation resale detection model based on the enhanced relationship graph convolutional network, specifically including: integrating the transaction context information mining module based on self-attention mechanism, the transaction temporal causal relationship module based on decaying time sequence gate, and the gated residual connection and relationship-specific weighted fusion module contained in the enhanced relationship graph convolutional network to form a single-layer structure of the enhanced relationship graph convolutional network, and stacking the single-layer structure in multiple layers to form a complete enhanced relationship graph convolutional network; adapting to the feature learning requirements of the multi-relationship transaction topology graph, deeply identifying transaction relationships, extracting transaction association features, and determining violations on the input multi-relationship transaction topology graph, and outputting the illegal resale situation existing in the transaction relationships; wherein, the relationship-aware attention weight output by the transaction context information mining module based on self-attention mechanism and the temporal gate weight output by the transaction temporal causal relationship module based on decaying time sequence gate are used to improve the gated residual connection and relationship-specific weighted fusion module.

[0009] According to one embodiment of the present invention, the improvement of the transaction context information mining module based on the self-attention mechanism specifically includes: separating basic transaction features and context features from the node features of each transaction, mapping the context features to an embedding space of a preset dimension; calculating the bilinear similarity of the mapped context features between adjacent transaction nodes to quantify the semantic consistency of the transaction content; concatenating the features of the central transaction node, the features of neighboring transaction nodes, and the bilinear similarity, calculating the attention score through a trainable attention vector and an activation function, and normalizing the attention score to obtain the relationship-aware attention weight.

[0010] According to one embodiment of the present invention, the improvement of the transaction temporal causal relationship module based on decaying time gate specifically includes: obtaining the transaction timestamps corresponding to adjacent transaction nodes, and calculating the transaction time difference between the adjacent transaction nodes; constructing a temporal direction control mechanism based on the transaction time difference, so that the direction of feature message transmission conforms to the temporal causal relationship of the order of transaction occurrence; constructing a time decay mechanism based on the transaction time difference, so that the adjacent transaction nodes with closer time intervals have higher association weights; and fusing the temporal direction control result and the time decay result to obtain the temporal gating weight.

[0011] According to one embodiment of the present invention, the improved gated residual connection and relation-specific weighted fusion module specifically includes: multiplying the relation-aware attention weights by the temporal gating weights to obtain message passing weights; configuring corresponding relation-specific weight matrices for role-sharing semantic edges and role-exchange semantic edges in the multi-relationship transaction topology graph; aggregating neighbor messages of adjacent transaction nodes based on the message passing weights and the relation-specific weight matrices to obtain aggregated neighbor messages; constructing a gated residual connection through a trainable gated scalar to fuse the transaction node's own features with the aggregated neighbor messages, and outputting the fused transaction node features.

[0012] According to one embodiment of the present invention, the method further includes training the data resale violation detection model based on the enhanced relational graph convolutional network, specifically including: inputting the multi-relationship transaction topology graph with resale violation labels as training samples into the data resale violation detection model; performing forward propagation through the data resale violation detection model to obtain resale violation prediction results of transaction nodes; calculating loss function values ​​based on the resale violation prediction results and the labels; updating the trainable parameters of the data resale violation detection model through backpropagation according to the loss function values; and completing model training until the loss function values ​​converge or a preset number of iterations is reached.

[0013] This invention also provides a data resale violation detection system based on an enhanced relational graph convolutional network, comprising: a transaction relationship construction module, used to acquire data transaction information from multiple platforms and construct a multi-relationship transaction topology graph with each transaction as a node and including role-sharing and role-exchange semantic edges; and a resale violation detection module, used to input the multi-relationship transaction topology graph into a data resale violation detection model based on an enhanced relational graph convolutional network to detect the existence of resale violations in the transaction information.

[0014] This invention constructs a multi-relationship transaction topology graph containing semantic edges of role sharing and role exchange, accurately depicting the complex relationships and role differences among subjects in multi-platform data transactions, thus solving the problem that traditional methods cannot adapt to multi-relationship transaction scenarios. Through targeted improvements to the three modules of the enhanced relationship graph convolutional network, it can deeply integrate transaction context semantic features and temporal causal features, combining relationship-specific weighting and gated residual connections to achieve efficient feature aggregation and learning, significantly improving the model's ability to extract and identify features of illegal resale behavior. Compared to existing technologies, this invention effectively reduces the false negative and false positive rates of illegal resale detection, achieving higher detection accuracy. It is adaptable to complex multi-platform transaction scenarios, realizing end-to-end automated illegal detection, providing efficient and reliable technical support for data transaction security supervision, and demonstrating stronger practicality and generalization. Attached Figure Description

[0015] 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.

[0016] Figure 1 This is a flowchart illustrating the data resale violation detection method based on enhanced relational graph convolutional networks provided by this invention.

[0017] Figure 2 This is a schematic diagram of the process for obtaining data transaction information from multiple platforms provided by the present invention.

[0018] Figure 3 This is a schematic diagram of the process for constructing a transaction topology graph with role-sharing semantic edges, provided by the present invention.

[0019] Figure 4 This is a schematic diagram of the process for constructing a transaction topology graph with role-swapping semantic edges provided by the present invention.

[0020] Figure 5 This is a flowchart illustrating the improvement of the transaction context information mining module based on the self-attention mechanism provided by the present invention.

[0021] Figure 6 This is a flowchart illustrating the improvement of the transaction timing causality module based on decay timing gates provided by the present invention.

[0022] Figure 7 This is a flowchart illustrating the improvement of the gated residual connection and relationship-specific weighted fusion module provided by the present invention.

[0023] Figure 8 This is a schematic diagram of the training process for a data resale violation detection model based on an enhanced relational graph convolutional network, provided by the present invention.

[0024] Figure 9 This is a block diagram of the data illegal resale detection system based on enhanced relational graph convolutional networks provided by the present invention.

[0025] Figure label: 100: Data Illegal Resale Detection System Based on Enhanced Relationship Graph Convolutional Network; 110: Transaction Relationship Construction Module; 120: Illegal Resale Detection Module. Detailed Implementation

[0026] 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.

[0027] The following is combined Figures 1 to 9 This invention describes a data resale violation detection method and system based on an enhanced relational graph convolutional network.

[0028] Figure 1 This is a flowchart illustrating the data violation resale detection method based on enhanced relational graph convolutional networks provided by this invention. Figure 1 As shown, the data resale violation detection method based on enhanced relational graph convolutional network provided by the present invention includes: in step S100, acquiring multi-platform data transaction information and constructing a multi-relational transaction topology graph with each transaction as a node and including role-sharing and role-exchange semantic edges; in step S200, inputting the multi-relational transaction topology graph into the data resale violation detection model based on enhanced relational graph convolutional network to detect the existence of resale violations in the transaction information.

[0029] Specifically, in step S100, firstly, tagged historical transaction data records from multiple data trading platforms within the target time period are acquired, and these standardized transaction data records are denoted as... This data contains all information for each transaction, including the index, identities of the buyer and seller, transaction characteristics, and violation flags. It serves as the original data source for constructing the subsequent multi-relationship transaction topology graph. Based on this, the construction of a multi-relationship transaction topology graph is carried out, ultimately generating an undirected heterogeneous graph form of the multi-relationship transaction topology graph, which is denoted as . This topology graph uses each transaction as a node, and the node set... Depend on The index of all transactions constitutes the node feature set. for Features of all transactions, node tag set for All illegal resale markers and relationship sets in the transaction. The edge set contains two types of semantic relationships: role sharing and role exchange. It is the main edge structure of the topological graph, consisting of a set of role-shared edges. And role-swapping edge set Together constitute, that is .

[0030] exist During the construction process, first from Extract buyer identity identifier set and seller identity identifiers The edge creation function defined in this invention Complete the generation of various edge sets, , Input the function respectively, and the results will be added to the initialized data in sequence. ,Will and The merged set is input into this function, and the result is added to the initialized set of mixed undirected edges. Then through set operations Generate a set of role-swapping edges, ultimately relying on the generated... and Determine the set of edge correspondences Complete the entire multi-relationship transaction topology diagram. The construction of this topology graph accurately depicts the complex relationships and role differences among entities in multi-platform transactions through two types of semantic edges. This provides a structured input for subsequent detection models that directly represent potential resale links, forming the foundation for accurate detection.

[0031] In step S200, the multi-relationship transaction topology diagram constructed in step S100 is... The complete input is fed into a data violation resale detection model built on an enhanced relational graph convolutional network. This detection model is an improvement and enhancement of the classic relational graph convolutional network. In this embodiment, a detection model containing two layers of enhanced relational graph convolutional network is constructed. The message passing mechanism introduces relation-aware attention weighting based on context similarity and decaying temporal gating based on causal order. Simultaneously, the gating residual connection dynamically balances the information ratio between the node's own features and the aggregated neighbor features. After receiving the topology graph input, the model first separates the basic transaction features and context features from the node features through a transaction context information mining module based on a self-attention mechanism, mapping the context features to... After embedding the nodes in the dimensional space, the semantic similarity of the transaction context between adjacent nodes is calculated using the bilinear similarity formula. Then, the node features and similarity are concatenated and normalized using an activation function and softmax to obtain the relation-aware attention weights. Subsequently, the transaction time-series causality module based on decaying time-series gates is used to obtain the transaction timestamps of adjacent nodes and calculate the time difference. The directional control weights are obtained through a time-series directional gate function and an exponential decay function, respectively. and time decay weight Multiplying the two yields the timing gating weights. Then multiply the attention weights by the temporal gating weights, and combine this with a relation-specific weight matrix. Aggregate neighbor node messages, through , The message passed to a single edge is obtained, and finally the gated residual connection formula is used. , The node features are updated. The model captures high-order correlation features in the transaction topology graph through forward propagation of two augmented network layers. Finally, the output features of the second layer are mapped to the [0,1] interval to obtain the predicted probability that each transaction node is involved in illegal resale. This allows for in-depth detection and analysis of illegal resale activities within transaction information.

[0032] The data resale violation detection model based on the enhanced relational graph convolutional network organizes and judges the detection analysis results, and the model outputs the predicted probability. Representative transaction node To determine the probability of illegal resale transactions, this embodiment uses a preset probability threshold. The model makes judgments, classifying transactions with predicted probabilities higher than a preset threshold as illegal resale transactions and those with probabilities lower than the threshold as normal transactions. Finally, the model aggregates all judgment results and outputs a complete transaction information report showing the illegal resale results, which presents a multi-relationship transaction topology graph. This includes identifying which transaction nodes constitute illegal resale transactions, the associated nodes and characteristic information of illegal transactions, and the chain structure of illegal resales, thus clarifying the scope and specific circumstances of illegal resales. The detection model in this embodiment is trained and optimized using a binary cross-entropy loss function, the formula of which is: , The Adam optimizer and initial learning rate of 0.001 were used to ensure the accuracy and reliability of the output results. This result provides an intuitive and accurate basis for data transaction security supervision. The entire data resale violation detection process based on reinforced graph convolutional networks is now complete. This process can be directly applied to new sets of transaction records with unknown labels. Simply repeat step S100 to construct the corresponding transaction topology graph, input it into the trained model, and the automatic identification of resale violation nodes can be achieved, demonstrating strong generalization ability and practicality.

[0033] Figure 2 This is a schematic diagram of the process for obtaining multi-platform data transaction information provided by the present invention. For example... Figure 2 As shown, according to an embodiment of the present invention, obtaining multi-platform data transaction information specifically includes: in step S111, parsing the obtained multi-platform data transaction information and extracting the transaction index, buyer identity identifier, and seller identity identifier corresponding to each transaction; in step S112, separating the buyer identity identifier set, seller identity identifier set, and mixed identity identifier set from the parsed transaction information; the mixed identity identifier set is the union of the buyer identity identifier set and the seller identity identifier set; in step S113, defining a side creation function, the input of which is the identity identifier set and the parsed transaction information, and the output is an undirected edge set; the execution logic of the side creation function is: traversing each identity identifier in the input identity identifier set, matching all transaction indices in which the identity identifier participates as the corresponding role in the transaction information, performing pairwise undirected connections on all transaction indices associated with the same identity identifier to generate undirected edges, and summarizing them into an edge set.

[0034] Specifically, in step S111, a comprehensive standardized parsing operation is performed on the acquired multi-platform data transaction information. The acquired transaction information covers the original transaction records of multiple data trading platforms. This type of data is multi-source and heterogeneous, lacking a unified structure and identification standard. It is necessary to extract the main and standardized identification information through parsing. The goal of the parsing operation is to accurately extract three types of identification information corresponding to each transaction from the massive amount of original multi-platform transaction information: a unique transaction index for each transaction, the buyer's identity identifier corresponding to the buyer entity participating in the transaction, and the seller's identity identifier corresponding to the seller entity participating in the transaction. Through this parsing operation, the multi-source and heterogeneous original transaction data is transformed into structured and standardized transaction data records. This data record serves as the sole unified data source for all subsequent identity set partitioning, edge creation function calls, and generation of various edge sets. The transaction index will serve as the unique identifier for nodes in the subsequent multi-relationship transaction topology graph. The buyer's identity identifier and the seller's identity identifier are the basis for subsequent role-based identity set partitioning and generation of different semantic edges. The standardized processing of the parsing operation directly determines the accuracy and uniformity of all subsequent topology graph construction-related operations.

[0035] In step S112, the standardized transaction data obtained in step S111 is recorded. The identification of the transaction entities is systematically classified and integrated, from Extract all occurrences of buyer identity identifiers and integrate them into a separate set of buyer identity identifiers. This collection contains the identity identifiers of all entities participating in transactions as buyers across multiple platforms; simultaneously from Extract all appearing seller identity identifiers and integrate them into a separate set of seller identity identifiers. This set contains the identity identifiers of all entities participating in transactions as sellers across multiple platforms. Based on this, [the following is further analyzed / analyzed]. and Perform a union operation to obtain a set of hybrid identity identifiers. Satisfying set operation relations , This collection encompasses all identity identifiers of entities participating in multi-platform transactions, with no duplicates or omissions, achieving complete integration of all transaction entities. This step completes the role-based classification of transaction entity identity identifiers, resulting in... , , The three sets of identity identifiers are the input parameters for subsequent calls to the edge creation function to generate different edge sets. Different sets of identity identifiers correspond to edge sets with different semantics, laying a solid foundation for the differentiation of edges with different semantics in the subsequent multi-relationship transaction topology graph.

[0036] In step S113, a dedicated edge creation function is defined to automate and standardize the generation of various edge sets in the future. , The input parameters, output results, and execution logic of the function are clearly defined. This function is generated subsequently. , The unified execution basis for undirected edge sets is also key to ensuring consistency in the rules for generating various edge sets. The function's input includes two parameters: one is any set of identity identifiers. , The value is obtained from the separation in step S112. , or The first is the standardized transaction data record obtained from the parsing in step S111, which can be flexibly called according to the needs of subsequent edge set generation; the second is the standardized transaction data record obtained from the parsing in step S111. This provides a unified and complete data source for the function's execution. The function's output is a set of undirected edges. This set contains all the sets of identities input. The undirected edges generated by connecting the transaction indexes of each entity in the process are the basic building blocks of various subsequent edge sets. This edge creation function has a fixed and clear execution logic; during function execution, it first initializes an empty set of undirected edges. This serves as the integration carrier for subsequently generating undirected edges; then, the input identity set is traversed. Each identity identifier in , for each Initialize a set of mappings from identity identifiers to transaction indices. Used to store All associated transaction indexes; then iterate through the standardized transaction data records. Accurately match the exist All transactions in which the corresponding trading role participates will have their corresponding transaction indices added to the mapping set. Complete a single After matching the transaction index, iterate through... All transaction indexes in the same All associated transaction indices are paired undirected joins. For each pair of paired joins, an undirected edge is generated and added to the initialization. Collection; waiting All identity identifiers After all iterations are completed, all generated undirected edges will be integrated. The collection is returned as the result of the function execution.

[0037] Overall, steps S111 to S113 are progressive and interconnected, completing the transformation from heterogeneous transaction data from multiple platforms to structured and standardized transaction data records. Then, a set of identity identifiers categorized by roles. , , Ultimately, this leads to the creation of the edge function with unified execution logic. The complete transformation process. All the data output from this step, along with the defined functions, is used to subsequently construct a set of role-shared edges. Mixed undirected edge set It forms the basis and direct execution basis, and is also the basis for subsequently generating role-swapping edge sets through set operations. The standardized processing method provides a unified standard and basis for the construction of the edge structure of the subsequent multi-relationship transaction topology graph. At the same time, this step is based solely on transaction behavior data and does not rely on any platform ownership information, which lays a key technical foundation for the present invention to achieve cross-platform data illegal resale detection.

[0038] Figure 3 This is a schematic diagram illustrating the process of constructing a transaction topology graph with role-sharing semantic edges, as provided by the present invention. For example... Figure 3 As shown, according to an embodiment of the present invention, constructing a transaction topology graph with role-shared semantic edges includes: in step S121, initializing an empty set of role-shared edges; in step S122, inputting a set of buyer identity identifiers into a creation edge function to obtain a subset of undirected edges shared by buyer roles, and adding the subset to the set of role-shared edges; in step S123, inputting a set of seller identity identifiers into a creation edge function to obtain a subset of undirected edges shared by seller roles, and adding the subset to the set of role-shared edges, thus completing the construction of the set of role-shared edges; all undirected edges in the set of role-shared edges are role-shared semantic edges in the multi-relationship transaction topology graph.

[0039] Specifically, in step S121, the initialization operation of the role-shared edge set is first performed to construct an empty role-shared edge set. This operation is the foundational preparation for constructing role-shared semantic edges. It provides a unified storage and integration carrier for the shared undirected edge subsets corresponding to the buyer and seller roles. Its initial state is an empty set without any undirected edges, ensuring that the subsequent addition of edge subsets for different roles can be carried out in an orderly manner. This makes the entire construction process of the role-shared edge set a systematic whole, avoiding the situation of scattered edge data and chaotic integration. It provides a standardized and unique target set for the edge subset integration operations in the subsequent two steps. In step S122, the edge subsets are integrated into the set. Figure 2 In step S112, the parsed transaction data is... The buyer identity set obtained from the separation Enter to Figure 2 The edge creation function defined in step S113 , At this point, the input parameters of the function for Input parameters To standardize the parsed transaction data, based on the predetermined input / output rules and execution logic of the edge function, a subset of shared undirected edges corresponding only to the buyer role is generated. The undirected edges within this subset are generated by pairwise undirected connections between transactions where the buyer identity serves as the corresponding role, accurately reflecting the association between transaction nodes under the buyer role. This subset of undirected edges shared by the buyer role is then further analyzed. Add to the set of shared edges initialized in step S121 In the middle, the first integration of shared undirected edges for buyer-side roles was completed, at this time... The shared undirected edges corresponding to all buyer roles have been included. In step S123, continue to... Figure 2 The set of seller identity identifiers obtained in step S112 Input to the edge creation function , Following the unified execution logic of this function, a subset of shared undirected edges corresponding only to the seller role is generated. The undirected edges within this subset are generated by pairwise undirected connections between transactions where the seller identity serves as the corresponding role, accurately reflecting the association between transaction nodes under the seller role. This subset of undirected edges shared by the seller role is then used to further refine the data. Add to the role-shared edge set of the integrated buyer edge subset In this context, the set of role-sharing edges... The integration of all undirected edges shared by the buyer and seller sides has been completed, forming the complete set of role-sharing edges. The construction operation is now complete. This final set of role-shared edges is now finished. The undirected edges contained therein constitute the multi-relationship transaction topology graph. The corresponding role-sharing semantic edges, as one of the two main edge structures in multi-relationship transaction topology graphs, accurately capture the relationships between transaction nodes formed when the same transaction entity participates in multiple transactions under the same transaction role. The completeness and accuracy of their construction directly determine the feature representation capability of the subsequent multi-relationship transaction topology graph, and also lay the foundation for subsequent hybrid undirected edge sets. Generation and role-swapping edge set The set operations provide crucial foundational data, while also offering precise topological support for detection models to mine the role association features of transaction entities and identify illegal resale behaviors.

[0040] Figure 4 This is a schematic diagram illustrating the process of constructing a transaction topology graph with role-exchange semantic edges, as provided by the present invention. For example... Figure 4As shown, according to an embodiment of the present invention, a transaction topology graph of role-exchange semantic edges is constructed, including: in step S131, an empty set of mixed undirected edges and an empty set of role-exchange edges are initialized; in step S132, a set of mixed identity identifiers is input into the edge creation function to obtain a set of mixed undirected edges associated with all transaction subjects; in step S133, a set operation is performed on the set of mixed undirected edges to remove all undirected edges in the set of role-shared edges, and the remaining undirected edges are summarized to form a set of role-exchange edges; all undirected edges in the set of role-exchange edges are role-exchange semantic edges in the multi-relationship transaction topology graph.

[0041] Specifically, in step S131, the initialization operations for the two types of edge sets are first performed, constructing empty mixed undirected edge sets respectively. Exchangeable edge set with empty roles .in, It serves as a temporary carrier for storing all the associated edges of the transaction entities, and is used to receive all the undirected edges generated by the mixed identity set; As the final target set to be constructed, it is initially empty, providing a dedicated storage medium for the results of subsequent set operations. This initialization operation ensures the orderliness of the subsequent edge set generation and filtering process, avoids confusion between different types of edge data, and provides a standardized set foundation for the logic from full generation to precise filtering.

[0042] In step S132, Figure 2 The hybrid identity set obtained in step S112 , Enter to Figure 2 The edge creation function defined in step S113 In the middle. At this time, the function's input parameters for Input parameters The data remains standardized after parsing, and iterates through the unified execution logic of this function. The system identifies all transaction entities and matches all transaction indices for each identity. It then performs pairwise undirected joins on transaction indices associated with the same identity, ultimately generating a mixed set of undirected edges containing all associated edges for all transaction entities. This set encompasses all undirected edges formed by transaction entities participating in transactions in any role (buyer or seller). It includes both previously generated role-sharing edges and role-exchange edges to be filtered, serving as the full data foundation for subsequent accurate separation of role-exchange edges.

[0043] In step S133, the mixed undirected edge set is processed by set operations. The process involves screening and purification, with the operational logic being... Remove from the total number of undirected edges included. Figure 3 The set of role-sharing edges constructed in the steps All undirected edges within the set, and the remaining undirected edges that were not removed, are combined to form a set of role-swapping edges. The characteristic of these residual undirected edges is that the connected transaction nodes correspond to the same transaction entity, but assume different transaction roles in different transactions (e.g., a buyer in one transaction and a seller in another), which precisely depicts the role exchange relationship of the transaction entity. This set operation achieves a precise separation between role-sharing edges and role-exchange edges, making the boundaries of the two types of semantic edges clear and unambiguous.

[0044] The final set of role-swapping edge sets All undirected edges in the graph constitute the multi-relational transaction topology. The corresponding role-exchange semantic edges. These semantic edges, together with role-sharing semantic edges, constitute the complete edge set of the topological graph. Both methods comprehensively depict the complex relationships between entities in multi-platform transactions from two dimensions: association within the same role and cross-role association. In particular, they provide crucial topological support for detection models to uncover hidden illegal resale chains such as buy-before-sell and multi-role collaboration. This is one of the technical features of this invention that adapts to multi-relationship transaction scenarios and improves the accuracy of violation detection. Thus, the multi-relationship transaction topology diagram is complete. The edge structure construction is now complete, laying a solid structured data foundation for feature learning and violation detection in the subsequent enhanced relational graph convolutional network model.

[0045] According to an embodiment of the present invention, the method further includes constructing a data violation resale detection model based on an enhanced relational graph convolutional network. Specifically, this includes: integrating a transaction context information mining module based on a self-attention mechanism, a transaction temporal causal relationship module based on decaying time-series gates, and a gated residual connection and relation-specific weighted fusion module contained in the enhanced relational graph convolutional network to form a single-layer structure of the enhanced relational graph convolutional network; stacking the single-layer structure in multiple layers to form a complete enhanced relational graph convolutional network; adapting to the feature learning requirements of multi-relationship transaction topology graphs, deeply identifying transaction relationships, extracting transaction association features, and determining violations in the input multi-relationship transaction topology graph, and outputting the violation resale situation existing in the transaction relationships; wherein, the relation-aware attention weights output by the transaction context information mining module based on a self-attention mechanism and the temporal gate weights output by the transaction temporal causal relationship module based on decaying time-series gates are used to improve the gated residual connection and relation-specific weighted fusion module.

[0046] Specifically, the model building process revolves around three key stages: module improvement, single-layer integration, and multi-layer stacking. Each stage is closely linked to ensure that the network has powerful feature extraction and violation detection capabilities.

[0047] First, targeted improvements are made to the three main modules of the Enhanced Relationship Graph Convolutional Network, with each improvement focusing on addressing the feature capture deficiencies of existing technologies in multi-relationship transaction scenarios. Specifically, the improvement to the transaction context information mining module based on a self-attention mechanism enables precise quantification of semantic consistency in transaction content and differentiated attention allocation. In this improvement process, basic transaction features and context features are first separated from the node features of each transaction. Basic transaction features reflect the inherent attributes of the transaction, while context features associate semantically relevant information such as data product type and industry domain. Then, an embedding mapping matrix is ​​used to map the context features to a preset-dimensional embedding space, achieving feature standardization and dimensional unification. Subsequently, the bilinear similarity of the mapped context features between adjacent transaction nodes is calculated, and the cross-correlation between features is learned through a bilinear weight matrix, accurately quantifying the degree of semantic consistency of the transaction content. Based on this, the features of the central transaction node, the features of neighboring transaction nodes, and bilinear similarity are concatenated. A nonlinear transformation is introduced through a trainable attention vector and an activation function to calculate the initial attention score. Then, the attention scores corresponding to the same type of edge are normalized by the softmax function to obtain the relation-aware attention weight. This weight can automatically assign higher importance to the connection of nodes with high context similarity, providing a basis for differentiation in subsequent message transmission.

[0048] Secondly, the improvement of the transaction temporal causality module based on decaying time-series gating focuses on modeling the impact of the causal order of transactions and time intervals on the strength of association. The improvement first obtains the transaction timestamps of adjacent transaction nodes and calculates the time difference between them, using this as the basis for temporal feature modeling. A temporal direction control mechanism is constructed based on the transaction time difference. Through trainable temporal direction penalty parameters and transaction interval parameters, message transmission from earlier to later transactions is rewarded, while reverse transmission and transactions with excessively short time intervals are penalized, ensuring that the direction of feature message transmission conforms to the temporal causal relationship of transaction occurrence. Simultaneously, a time decay mechanism is constructed based on the transaction time difference. An exponential decay function ensures that adjacent transaction nodes with closer time intervals have higher association weights, while the weight decays more significantly with longer time intervals, accurately characterizing the time urgency of resale behavior. Finally, the temporal direction control result is multiplied by the time decay result to obtain the temporal gating weights, achieving a fusion model of temporal causality and time decay effects.

[0049] Furthermore, the improved gated residual connection and relation-specific weighted fusion module integrates the output weights of the first two modules and combines them with the differentiated features of multi-relationship edges to achieve efficient neighbor message aggregation and node feature updates. The key to this improvement lies in using the relation-aware attention weights output by the transaction context information mining module based on a self-attention mechanism, and the temporal gating weights output by the transaction temporal causality module based on decaying temporal gates. Multiplying these two weights yields the message passing weights, reflecting both the influence of contextual semantic similarity and incorporating temporal causal constraints. For role-sharing semantic edges and role-exchange semantic edges in the multi-relationship transaction topology graph, corresponding relation-specific weight matrices are configured, and the differentiated feature representations of the two types of semantic edges are learned through different weight matrices. Based on the message passing weights and relation-specific weight matrices, the neighbor messages of adjacent transaction nodes are weighted and aggregated, fusing the features of different neighbor nodes according to their weights to obtain the aggregated neighbor messages. To avoid the loss of the inherent features of nodes during feature fusion, a trainable gated scalar is introduced to construct a gated residual connection, which dynamically balances the information ratio between the transaction node's own features and the aggregated neighbor messages. The fused transaction node features are output through an activation function, which not only retains the main features of the node itself, but also integrates the association information of neighbor nodes, thereby improving the completeness and effectiveness of feature representation.

[0050] After improving the three main modules, they are integrated into a single-layer structure of an enhanced relational graph convolutional network. In this single-layer structure, the three modules work collaboratively in the logical order of contextual feature mining, temporal causal modeling, and message aggregation and fusion: first, the transaction context information mining module obtains relation-aware attention weights; then, the transaction temporal causal relationship module obtains temporal gating weights; finally, the gated residual connection and relation-specific weighted fusion module completes neighbor message aggregation and node feature updates, forming a complete single-layer feature learning process. To further enhance the network's ability to capture high-order relational features in the transaction topology graph, this single-layer structure is stacked in multiple layers to form a complete enhanced relational graph convolutional network. This multi-layered network structure can gradually mine multi-hop relationships between transaction nodes through multiple rounds of feature propagation and aggregation, accurately capturing complex illegal resale link features, allowing the network to deeply adapt to the feature learning needs of multi-relationship transaction topology graphs.

[0051] The final data resale violation detection model processes the input multi-relationship transaction topology graph through a complete augmented relational graph convolutional network: the network first updates the features of each transaction node in multiple rounds, deeply identifies the semantic associations, temporal constraints and multi-type edge features in the transaction relationship, and extracts transaction association features that can accurately represent illegal resale behavior; then the node features output from the final layer are input into a fully connected classifier, and the output is mapped to the [0,1] interval through an activation function to obtain the predicted probability of each transaction node being an illegal resale; finally, the violation is judged based on the predicted probability and a preset threshold, and the illegal resale situation in the transaction relationship is output, realizing end-to-end detection from feature learning to violation judgment, effectively improving the ability to identify cross-platform and covert illegal resale behavior.

[0052] Figure 5 This is a flowchart illustrating the improved transaction context information mining module based on the self-attention mechanism provided by the present invention. Figure 5 As shown, according to an embodiment of the present invention, the transaction context information mining module based on the self-attention mechanism is improved, specifically including: in step S211, separating the basic transaction features and context features from the node features of each transaction, and mapping the context features to an embedding space of a preset dimension; in step S212, calculating the bilinear similarity of the mapped context features between adjacent transaction nodes to quantify the semantic consistency of the transaction content; in step S213, concatenating the features of the central transaction node, the features of neighboring transaction nodes, and the bilinear similarity, calculating the attention score through a trainable attention vector and activation function, and normalizing the attention score to obtain the relationship-aware attention weight.

[0053] Specifically, in step S211, the refined separation and Z-dimensional embedding mapping of transaction node features are completed, providing a standardized feature carrier for semantic similarity calculation. First, for each neighbor node j connected by a relation type r edge in the l-th layer network, its node features... It will be broken down into two independent sub-features: one is the basic transaction feature. Another type is transaction context information features. The splitting relationship satisfies the formula: , Where ⨁ represents the feature concatenation operation, ( The total dimension of node features. (For contextual features), it mainly includes inherent attributes such as transaction amount and transaction number that are unrelated to semantic association; Focusing on semantic information that reflects the relationship between transaction content, such as data product type, industry sector, and data usage, is a characteristic for determining whether a transaction has a resale relationship.

[0054] Subsequently, through a trainable embedding mapping matrix ( (This is a pre-defined embedding space dimension). To balance the complexity of multi-platform transaction features with model computational efficiency, Z-dimensionality is set to 128 dimensions, with its value adjustable flexibly between 32 and 256 dimensions. Lower dimensions (32 / 64) are suitable for small amounts of transaction data on a single platform, while higher dimensions (256) are suitable for massive transaction scenarios across multiple platforms. 128 dimensions represent the optimal choice, balancing semantic integrity and computational efficiency. The separated transaction context information features... ( Mapped to (central node) In the dimensional embedding space, we obtain a context feature vector after mapping to a unified dimension. The mapping formula is: , The purpose of this operation is to eliminate the high-dimensional heterogeneity of the original contextual features, transforming them into low-dimensional, dense vector representations that fully preserve semantic information while also leveraging trainable vectors. By learning the semantic association patterns of contextual features, a unified feature foundation is provided for subsequent similarity calculations between adjacent nodes, while improving computational efficiency.

[0055] In step S212, the focus is on accurately quantifying the semantic consistency of the transaction context between adjacent nodes through bilinear similarity calculation. This is based on the central node obtained in step S211. post-mapping context features and the context features after mapping of neighbor node j The semantic association strength between the two is calculated using a bilinear similarity algorithm, and the formula is defined as follows: , in, This indicates the L2 normalization operation, used to eliminate the interference of vector magnitude differences on similarity results; It is a trainable bilinear weight matrix that can adaptively learn two... Compared to traditional similarity calculation methods, the cross-correlation and higher-order interaction information between contextual feature vectors can more accurately capture the semantic relationships of transaction content. For example, the contextual features of the same type of data product in different transactions can be obtained through... The system learns to achieve higher similarity values. The final bilinear similarity value directly reflects the semantic consistency between the two transactions. The higher the similarity, the more similar the semantic information involved, and the more likely there is an illegal resale connection.

[0056] In step S213, attention scores are calculated by fusing multi-dimensional information, and relation-aware attention weights are obtained through normalization. First, a feature concatenation operation is performed: the center node... The Layer features Trainable weight matrix The result after transformation Neighbor nodes The Layer features Through the same The result after transformation and the bilinear similarity calculated in step S212 The features are then concatenated to form a comprehensive feature vector.

[0057] Subsequently, this synthesized feature vector is input into a trainable attention vector. ( To determine the type of edge (i.e., to configure a dedicated attention vector for each relation type, reflecting a "relationship-specific" attention mechanism), nonlinearity is introduced through linear transformation and the LeakyReLU activation function to calculate the initial attention score. The formula is: , in, Representing vector concatenation operations, the LeakyReLU activation function can effectively avoid the gradient vanishing problem while enhancing the model's ability to capture complex semantic relationships.

[0058] Finally, regarding the central node In relation The set of all neighbor nodes under the type edge The initial attention scores for each neighbor node are obtained using the softmax function. Normalization is performed to ensure that the sum of the attention weights of all neighboring nodes is 1, ultimately yielding the relation-aware attention weights. The formula is: , And the constraints are satisfied. .

[0059] This weighting automatically assigns higher importance to neighboring nodes with high semantic similarity, so that in subsequent message transmission, transaction nodes with close semantic relationships have a stronger impact on the feature updates of the central node, thereby accurately capturing the semantic relationship features of transaction content in illegal resale behavior.

[0060] Overall, the improved process from steps S211 to S213 fully replicates the technical details of this invention. From feature separation and Z-dimensional embedding mapping to bilinear similarity quantification, and then to attention score calculation and normalization, the formula definitions, parameter meanings, and operational logic of each step strictly adhere to the technical solution requirements, without any omissions. The improved module can accurately mine semantic association features in multi-relationship transactions, providing a differentiated attention allocation basis for enhancing relational graph convolutional networks. This allows the model to focus on key semantic associations in multi-platform and complex transaction scenarios, providing technical support for the accurate identification of illegal resale behavior.

[0061] Figure 6 This is a flowchart illustrating the improvement to the transaction timing causality module based on decaying timing gates provided by this invention. Figure 6 As shown, according to an embodiment of the present invention, the transaction time-series causal relationship module based on decaying time-series gates is improved, specifically including: in step S221, obtaining the transaction timestamps corresponding to adjacent transaction nodes and calculating the transaction time difference between adjacent transaction nodes; in step S222, constructing a time-series direction control mechanism based on the transaction time difference, so that the direction of feature message transmission conforms to the time-series causal relationship of the transaction occurrence; in step S223, constructing a time decay mechanism based on the transaction time difference, so that the adjacent transaction nodes with closer time intervals correspond to higher association weights; in step S224, fusing the time-series direction control result and the time decay result to obtain the time-series gate control weight.

[0062] Specifically, in step S221, the timestamps of adjacent transaction nodes are obtained and the transaction time difference is calculated, providing basic time-series characteristics for the subsequent construction of the time-series mechanism. Regarding the first... In a layered network, through relationships The central transaction node connected by the type edge Transaction nodes with neighbors Extract the corresponding transaction timestamps for each. and This timestamp is the actual time the transaction occurred and is the raw data characterizing the temporal relationship of transactions. The time difference between two nodes is calculated based on the extracted timestamps. The calculation formula is: ， Among them, time difference The sign of the value directly reflects the order in which the two transactions occurred: Representing neighboring nodes The corresponding transaction occurs at the central node. Before the corresponding transaction, This indicates the neighbor node. The corresponding transaction occurs at the central node. The corresponding transaction; time difference absolute value This reflects the actual time interval between the two transactions. This time difference is the fundamental feature for subsequently building the timing direction control mechanism and the time decay mechanism.

[0063] In step S222, a temporal direction control mechanism is constructed based on the transaction time difference, and the direction control weights are calculated to ensure that the direction of feature message transmission conforms to the temporal causal relationship of transaction occurrence. This step constructs a temporal direction gate through a trainable parameterized function, rewards message transmission that conforms to the causal order of past transactions to future transactions, and penalizes reverse transmission and transactions with excessively short time intervals. Finally, the results are mapped to the [0,1] interval through an activation function to obtain the direction control weights. The calculation formula is: , in, The sigmoid activation function is used to normalize the function calculation result to the [0,1] interval to ensure the reasonableness of the weight values; For trainable temporal orientation penalty parameters, which are related to ΔT The product term is used to implement rewards and penalties in the temporal direction: when ΔT>0 When a message is transmitted from the earliest transaction to the latest transaction, the product term is positive, increasing the directional control weight and resulting in a reward; when ΔT<0 When a message is passed from a later transaction to an earlier transaction, the product term is negative, reducing the directional control weight and thus incurring a penalty. The time difference penalty parameter is a trainable parameter that applies a global penalty to the transaction time difference, balancing the reward and penalty magnitudes for the time series direction. The trainable transaction interval parameter is used to learn the normal minimum transaction interval in multi-platform trading. Its negative term can penalize abnormal transaction behavior with excessively short time intervals, conforming to the time characteristics of illegal resale behavior. The final direction control weight is obtained. The closer the value is to 1, the more the message transmission direction of that edge conforms to the temporal causal relationship of the transaction; conversely, the less it conforms, the less it does.

[0064] In step S223, a time decay mechanism is constructed based on the transaction time difference, and the time decay weight is calculated to characterize the time-series feature that adjacent transaction nodes with closer time intervals correspond to higher association weights. This step uses an exponential decay function to decay the association weights based on the absolute value of the transaction time difference, accurately adapting to the time urgency characteristic of illegal resale behavior, which often involves completing transaction flows within a short period of time, thus obtaining the time decay weights. The calculation formula is: , in, This is the absolute value of the transaction time difference, representing the actual time interval between two transactions; For edge type The relevant trainable decay parameters can adaptively learn the corresponding decay rate based on the temporal characteristics of different semantic edges, such as role-sharing and role-exchange, thereby improving the model's adaptability to multi-relationship transaction scenarios. The mathematical properties of the exponential function determine the time interval between two transactions. The smaller the value, the greater the time decay weight. The closer the value is to 1, the higher the temporal correlation between the two; the time interval The larger the value, the closer the time decay weight is to 0, which means that the temporal correlation between the two is lower, thus realizing differentiated weight allocation based on time interval.

[0065] In step S224, the timing direction control result and the time decay result are fused to obtain the final timing gating weight, providing a timing factor with both causal order constraints and time decay constraints for subsequent message passing. This step uses the direction control weight obtained in step S232. The time decay weight obtained in step S233 Element-wise multiplication is performed to achieve deep fusion of the two temporal constraint features, resulting in temporal gating weights. The calculation formula is: , This timing gating weight Simultaneously, it integrates both temporal causal order and time interval decay features. Its numerical value depends on whether the message transmission direction conforms to the chronological order of transactions, and also on the proximity of the time intervals between two transactions. Only adjacent nodes that conform to both the causal transmission order of past transactions to future transactions and have close transaction time intervals can obtain a high temporal gating weight. This weight will serve as a temporal constraint factor in the subsequent message transmission process, allowing the model to focus more on the association of transaction nodes whose temporal features match illegal resale behavior when aggregating neighbor node features, providing crucial temporal dimension feature support for accurately identifying hidden illegal resale behavior.

[0066] Figure 7 This is a schematic diagram illustrating the improved gated residual connection and relationship-specific weighted fusion module provided by the present invention. For example... Figure 7As shown, according to an embodiment of the present invention, the improved gated residual connection and relation-specific weighted fusion module specifically includes: in step S231, multiplying the relation-aware attention weights and the temporal gating weights to obtain message passing weights; in step S232, configuring corresponding relation-specific weight matrices for role-sharing semantic edges and role-exchange semantic edges in the multi-relationship transaction topology graph; in step S233, aggregating neighbor messages of adjacent transaction nodes based on the message passing weights and relation-specific weight matrices to obtain aggregated neighbor messages; in step S234, constructing a gated residual connection through a trainable gated scalar, fusing the transaction node's own features with the aggregated neighbor messages, and outputting the fused transaction node features.

[0067] Specifically, in step S231, the two types of weights are fused to obtain the message passing weight, achieving a deep binding between contextual semantic constraints and temporal causal constraints. The input to this step is the output of the first two major modules: one is the relation-aware attention weight obtained from the transaction context information mining module based on the self-attention mechanism. The first weight characterizes the contextual semantic similarity association between adjacent transaction nodes; the second is the time-gated weight obtained based on the transaction time-series causal relationship module with decaying time-series gates. This weight integrates the temporal causal order of transactions with time decay characteristics. The message passing weight is obtained by fusing the two types of weights through element-wise multiplication. The calculation formula is: , in, Represents the edge type (role sharing or role exchange) in a multi-relationship transaction topology graph. As the central transaction node, As the central node In relation The neighboring transaction nodes are integrated. This fusion operation allows message passing weights to have both "semantic similarity priority" and "temporal rationality constraint": neighboring nodes with close semantic associations and temporal characteristics that conform to illegal resale behavior (buying before selling, tight timing) will receive higher message passing weights, providing a differentiated basis for subsequent neighbor message aggregation and ensuring that the model focuses on key related nodes.

[0068] In step S232, differentiated relation-specific weight matrices are configured for different types of semantic edges to adapt to the feature representation requirements of multi-relationship transactions. Since the semantic connotations of role-sharing semantic edges and role-exchange semantic edges are fundamentally different—role-sharing edges reflect associations under the same transaction role, such as multiple transactions by the same buyer, while role-exchange edges reflect resale associations across roles, such as transactions where the same entity buys first and then sells—the feature information carried by these two types of edges needs to be learned specifically through differentiated matrices. Therefore, dedicated relation-specific weight matrices are configured for each type of semantic edge: for role-sharing semantic edges, a weight matrix is ​​configured... ( For semantic edges involving role exchange, configure a weight matrix. Both types of matrices have the same dimension. , , where is the node feature dimension, and all are trainable parameters. During model training, it will adaptively learn the feature transformation rules of the two types of semantic edges, so as to achieve accurate extraction and expression of features of different relation types and avoid expression distortion caused by confusion of multiple relation features.

[0069] In step S233, based on the message passing weight and relationship-specific weight matrix, the weighted aggregation of neighbor messages is completed, integrating the association features of multiple neighbors and multiple relationships. The aggregation process proceeds logically as follows: grouping by relationship type - weighting neighbor by neighbor - summarizing within the same relationship: first, for the central node... The Layer features traverse its relationships across all relation types The set of neighboring nodes For each neighbor node First, through the correspondence weight matrix Features of neighboring nodes Perform feature transformation, then multiply by message passing weight. Receive the weighted message of a single edge The calculation formula is: , Subsequently, for the same relation type The weighted messages of all neighboring nodes are summed, and then the summation results of all relation types are accumulated to obtain the central node. Aggregated neighbor messages The overall aggregation formula is: , This polymerization process is both through Giving key neighbors a higher contribution rating, and through The feature differences that distinguish different relation types ultimately lead to It fully integrates the association features of multiple neighbors and multiple relationships, providing rich association information support for node feature updates.

[0070] In step S234, a gated residual connection is constructed to balance the information ratio between the node's own features and the aggregated neighbor messages, outputting the fused node features. To avoid the loss of the inherent features of the central node, such as the basic attributes and inherent semantics of transactions, during feature aggregation, and to dynamically adjust the fusion ratio of associated features, a trainable gated scalar is introduced. First, construct gated residual connections. Then, use gated scalar pairs to aggregate neighbor messages. Weighting is applied, while also considering the characteristics of the central node itself. Through the basic weight matrix (dimension) After feature transformation (which can be trained), use Weighting is applied to achieve a dynamic balance between the two types of features; then, the weighted features are summed, and a nonlinear transformation is introduced through the sigmoid activation function to output the center node. No. Features after layer fusion The calculation formula is: , The sigmoid activation function maps the fusion result to the [0,1] interval, ensuring the rationality of feature values; the gate scalar Adaptive adjustment through model training, when When the value approaches 1, the model focuses more on aggregating the association features of neighbors; when... When the value approaches 0, the model retains more of the inherent characteristics of the nodes, thereby dynamically balancing the integrity and relevance of the features and improving the expressive power of the node features.

[0071] Figure 8 This is a schematic diagram illustrating the training process of a data resale violation detection model based on an enhanced relational graph convolutional network, as provided by this invention. Figure 8As shown, according to an embodiment of the present invention, the method further includes training a data resale violation detection model based on an enhanced relational graph convolutional network, specifically including: in step S241, inputting a multi-relationship transaction topology graph with resale violation labels as training samples into the data resale violation detection model; in step S242, performing forward propagation through the data resale violation detection model to obtain the resale violation prediction results of the transaction nodes; in step S243, calculating the loss function value based on the resale violation prediction results and the labels; in step S244, updating the trainable parameters of the data resale violation detection model through backpropagation according to the loss function value; and in step S245, completing the model training until the loss function value converges or reaches a preset number of iterations.

[0072] Specifically, in step S241, the preparation and input of training samples are completed, providing a standardized data source for model training. The training samples are multi-relationship transaction topology graphs with labels indicating illegal resale. The construction process of this topology map is consistent with that described above: using labeled historical transaction data from multiple platforms within the target time period. Based on this, transaction indexes and the identities of buyers and sellers are extracted to construct a role-sharing edge set. Edge set with role swapping To form a complete set of edges , where the set of nodes For transaction index set, node feature set A set of node labels that combine the basic and contextual features of a transaction. For illegal resale of tags, Indicates transaction This is illegal resale. Indicates transaction This is for normal transactions. During the input process, it is necessary to ensure that the training samples have undergone data preprocessing, that the identities of both parties in the transaction are uniquely mapped through a unified identity authentication platform, and that the structure and feature dimensions of the topology graph are consistent to avoid affecting the training effect due to data heterogeneity. This standardized training sample is then completely input into the data resale detection model based on an enhanced relational graph convolutional network to be trained. The model will then use this sample as a basis for feature learning and parameter optimization.

[0073] In step S242, the prediction results of illegal resale of transaction nodes are output through the model forward propagation process. The forward propagation process strictly follows the hierarchical computation logic of the enhanced relation graph convolutional network. Taking the two-layer enhanced network constructed in this embodiment as an example, the specific process is as follows: First, the multi-relationship transaction topology graph in the training samples... After being input into the model, the first augmentation network layer calculates the relationship-aware attention weights through a transaction context information mining module based on a self-attention mechanism. Then, through the transaction time-series causality module based on decaying time-series gates, the time-series gating weights are obtained. Subsequently, the message passing weight is obtained by multiplying the dual weights through the gated residual connection and relationship-specific weighted fusion module, and then combined with the relationship-specific weight matrix. Aggregate neighbor messages and update node features using the gated residual connection formula. The node features output from the first layer of the network will be used as input to the second layer. This feature update process is repeated until the high-order node features output from the second layer are obtained. The high-order node features are input into a fully connected classifier for linear transformation, and then the output is mapped to the [0,1] interval using a Sigmoid activation function to obtain each transaction node. Probability of illegal resale ,Right now: , in, This is the trainable weight matrix for the fully connected layer. For bias terms, The magnitude of the value directly reflects the transaction. The closer the value is to 1, the higher the likelihood of illegal resale.

[0074] In step S243, based on the prediction results obtained from forward propagation and the sample labels, the loss function value is calculated to quantify the model prediction error. This invention uses binary cross-entropy (BCE) as the loss function, which is suitable for binary classification tasks and can effectively measure the difference between the predicted probability and the true label. The calculation formula is as follows: , Where N is the total number of transaction nodes in the training samples, that is, the number of nodes in the topology graph G. For transaction nodes The actual label (0 or 1). This represents the corresponding predicted probability. (Loss function value) The larger the value, the greater the deviation between the model's predictions and the actual situation; conversely, the smaller the deviation, the better the model's fit. This loss value will serve as the basis for subsequent backpropagation parameter updates, guiding the model to optimize in the direction of reducing prediction errors.

[0075] In step S244, all trainable parameters of the model are updated based on the loss function value using the backpropagation algorithm, thereby achieving iterative optimization of the model. First, the loss function value... To achieve this, the chain rule is used to calculate the gradient of the loss with respect to all trainable parameters in the model, i.e., the degree of influence of parameter changes on the loss. The trainable parameters involved include: the embedding mapping matrix. Bilinear weight matrix Attention vector Characteristic transformation matrix Timing direction penalty parameters Time difference penalty parameters Transaction interval parameters Attenuation parameters Relationship-specific weight matrix Gated scalar Fully connected layer weight matrix With bias term Then, the Adam optimizer is used with an initial learning rate of 0.001. All trainable parameters are adaptively updated based on the calculated gradients. The update logic is as follows: , in, This represents any trainable parameter in the model. The learning rate is 0.001. This represents the gradient of the loss function with respect to this parameter. Through this update process, the model will gradually adjust the parameter values, reduce the loss function value, and improve its ability to capture features related to illegal resale.

[0076] In step S245, it is determined whether training should terminate to ensure that the model reaches a stable optimal state. The training process will repeat the forward propagation-loss calculation-backward propagation update process from steps S242 to S24 until one of the following two termination conditions is met: one is the loss function value. Convergence, meaning that within a preset number of rounds (e.g., 10 rounds), the change in the loss function value is less than a preset threshold (e.g., 10 rounds). -5 The first termination condition indicates that the model has fully learned the feature patterns in the training samples, and further training will not significantly improve performance. The second termination condition is that the training rounds have reached the preset number of iterations (100 in this embodiment) to avoid overfitting due to excessive training, meaning the model only adapts to the training samples and its generalization ability to new samples decreases. When any termination condition is met, the model training is complete. At this point, all trainable parameters of the model will be fixed, providing a stable ability to detect illegal resales. This allows the model to be directly used to process new, unlabeled transaction topologies, enabling automatic identification of illegal resale nodes.

[0077] Overall, the training process from steps S241 to S245 forms a complete closed loop of data input, model learning, parameter optimization, and performance stabilization. Through this training process, the model can fully explore the semantic association features, temporal causal features, and multi-type edge features in multi-relationship transaction topologies, accurately learn the patterns of illegal resale behavior, and provide high-precision, high-generalization model support for subsequent violation detection.

[0078] Figure 9 This is a block diagram of the data illegal resale detection system based on enhanced relational graph convolutional networks provided by this invention. Figure 9 As shown, the present invention also provides a data illegal resale detection system 100 based on an enhanced relational graph convolutional network, comprising: a transaction relationship construction module 110, used to acquire data transaction information from multiple platforms and construct a multi-relationship transaction topology graph with each transaction as a node and including role-sharing and role-exchange semantic edges; and an illegal resale detection module 120, used to input the multi-relationship transaction topology graph into a data illegal resale detection model based on an enhanced relational graph convolutional network to detect illegal resale in the transaction information.

[0079] The data resale violation detection system based on enhanced relational graph convolutional networks provided by this invention corresponds one-to-one with the aforementioned detection methods. The functional modules of the system and the steps of the methods are precisely matched and can be mutually referenced. The specific correspondence is as follows: The functional implementation of the transaction relationship construction module is completely consistent with the process of obtaining multi-platform data transaction information and constructing a multi-relationship transaction topology graph in the method. Its operational logic, such as data parsing, identity identifier set separation, edge creation function call, and construction of two types of semantic edges, all follow the detailed description in the method section. The detection logic of the illegal resale detection module corresponds one-to-one with the process of inputting the multi-relationship transaction topology graph into the enhanced relationship graph convolutional network model for illegal detection in the method. The implementation details of the model's module composition, feature extraction, weight calculation, message aggregation, and illegal judgment are completely synchronized with the technical solutions such as model construction, module improvement, and training process in the method section.

[0080] Since the conception, technical principles, and implementation steps of the system have been described in detail in the detection method section, in order to avoid repetition, the system section will not elaborate on the relevant technical details. Those skilled in the art can directly refer to the specific implementation methods of the aforementioned detection methods to understand and implement the detection system.

[0081] This invention provides a data resale violation detection method based on enhanced relational graph convolutional networks. Through precise construction of multi-relational transaction topology graphs, targeted improvements to three major modules, systematic model training, and standardized data preprocessing, it forms an end-to-end detection solution adaptable to complex transaction scenarios across multiple platforms. This method overcomes the limitations of traditional detection methods, such as insufficient modeling of transaction relationship structures and a lack of temporal and semantic feature capture. Furthermore, by organically integrating relation-aware attention mechanisms, decaying temporal gating, and gated residual connections, it achieves in-depth mining of multi-dimensional features of resale violations, significantly improving detection accuracy and generalization ability. Its independence from single-platform ownership information enables it to effectively address cross-platform, covert resale violations, providing reliable technical support for the security supervision of the data transaction market and possessing broad practical application prospects and promotional value.

[0082] 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 data violation resale detection method based on an enhanced relational graph convolutional network, characterized in that, include: Acquire data transaction information from multiple platforms and construct a multi-relational transaction topology graph with each transaction as a node and including semantic edges of role sharing and role exchange. The multi-relationship transaction topology graph is input into a data violation resale detection model based on an enhanced relational graph convolutional network to detect any violations in the transaction information.

2. The method according to claim 1, characterized in that, The acquisition of multi-platform data transaction information specifically includes: The acquired multi-platform data transaction information is analyzed to extract the transaction index, buyer identity identifier, and seller identity identifier corresponding to each transaction. From the parsed transaction information, a set of buyer identity identifiers, a set of seller identity identifiers, and a set of mixed identity identifiers are obtained; the set of mixed identity identifiers is the union of the set of buyer identity identifiers and the set of seller identity identifiers. Define an edge creation function. The input of the function is a set of identity identifiers and the parsed transaction information. The output is a set of undirected edges. The execution logic of the edge creation function is as follows: traverse each identity identifier in the input set of identity identifiers, match all transaction indices in which the identity identifier participates as the corresponding role in the transaction information, perform pairwise undirected connections on all transaction indices associated with the same identity identifier to generate undirected edges, and summarize them into an edge set.

3. The method according to claim 2, characterized in that, Construct a transaction topology graph with role-shared semantic edges, including: Initialize an empty set of shared edges for roles; Input the set of buyer identity identifiers into the edge creation function to obtain a subset of undirected edges shared by buyer roles, and add the subset to the set of role-shared edges. The seller identity set is input into the edge creation function to obtain a subset of undirected edges shared by seller roles. This subset is then added to the set of role-shared edges to complete the construction of the set of role-shared edges. All undirected edges in the set of role-shared edges are role-shared semantic edges in the multi-relationship transaction topology graph.

4. The method according to claim 3, characterized in that, Construct a transaction topology graph for role-swapping semantic edges, including: Initialize an empty set of mixed undirected edges and an empty set of role-swapping edges; Input the set of hybrid identity identifiers into the edge creation function to obtain a set of hybrid undirected edges associated with all transaction entities; A set operation is performed on the mixed undirected edge set to remove all undirected edges from the set of role-shared edges, and the remaining undirected edges are aggregated to form a set of role-exchange edges; all undirected edges in the set of role-exchange edges are role-exchange semantic edges in the multi-relationship transaction topology graph.

5. The method according to claim 1, characterized in that, The method also includes constructing the data illegal resale detection model based on the enhanced relational graph convolutional network, specifically including: The transaction context information mining module based on self-attention mechanism, the transaction temporal causal relationship module based on decaying time sequence gate, and the gated residual connection and relation-specific weighted fusion module contained in the enhanced relation graph convolutional network are integrated to form a single-layer structure of the enhanced relation graph convolutional network, and the single-layer structure is stacked in multiple layers to form the complete enhanced relation graph convolutional network. To adapt to the feature learning requirements of the multi-relationship transaction topology graph, the system deeply identifies transaction relationships, extracts transaction association features, and determines violations in the input multi-relationship transaction topology graph, outputting the illegal resale situation in the transaction relationships; Specifically, the relation-aware attention weights output by the transaction context information mining module based on the self-attention mechanism and the temporal gating weights output by the transaction temporal causal relationship module based on decaying temporal gates are used to improve the gated residual connection and relation-specific weighted fusion module.

6. The method according to claim 5, characterized in that, The improvements to the transaction context information mining module based on the self-attention mechanism specifically include: The basic transaction features and context features are separated from the node features of each transaction, and the context features are mapped to the embedding space of a preset dimension. Calculate the bilinear similarity of contextual features mapped between adjacent transaction nodes to quantify the semantic consistency of transaction content; The features of the central transaction node, the features of neighboring transaction nodes, and bilinear similarity are concatenated. An attention score is calculated using a trainable attention vector and an activation function. The attention score is then normalized to obtain the relationship-aware attention weight.

7. The method according to claim 5, characterized in that, The improvement to the transaction time-series causal relationship module based on decay time-series gates specifically includes: Obtain the transaction timestamps corresponding to adjacent transaction nodes, and calculate the transaction time difference between the adjacent transaction nodes; Based on the transaction time difference, a timing direction control mechanism is constructed to ensure that the direction of feature message transmission conforms to the causal relationship of the order of transaction occurrence. Based on the transaction time difference, a time decay mechanism is constructed so that the adjacent transaction nodes with closer time intervals have higher association weights. The timing gating weights are obtained by fusing the timing direction control results and the time decay results.

8. The method according to claim 5, characterized in that, The improved gated residual connection and relation-specific weighted fusion module specifically includes: Multiply the relation-aware attention weight by the temporal gating weight to obtain the message passing weight; For the role-sharing semantic edges and role-exchange semantic edges in the multi-relationship transaction topology graph, respectively configure corresponding relationship-specific weight matrices; Based on the message passing weight and the relationship-specific weight matrix, the neighbor messages of adjacent transaction nodes are aggregated to obtain aggregated neighbor messages. By using a trainable gated scalar, a gated residual connection is constructed, which fuses the characteristics of the transaction node itself with the aggregated neighbor messages, and outputs the fused characteristics of the transaction node.

9. The method according to claim 1, characterized in that, It also includes training the data resale violation detection model based on the enhanced relational graph convolutional network, specifically including: The multi-relationship transaction topology graph with the label of illegal resale is used as a training sample and input into the data illegal resale detection model; The illegal resale prediction results of the transaction nodes are obtained by forward propagation through the aforementioned data illegal resale detection model; The loss function value is calculated based on the predicted results of illegal resale and the labeled information. Based on the loss function value, the trainable parameters of the data illegal resale detection model are updated through backpropagation; The model training is completed when the loss function value converges or the preset number of iterations is reached.

10. A data resale violation detection system based on an enhanced relational graph convolutional network, characterized in that, include: The transaction relationship construction module is used to obtain data transaction information from multiple platforms and construct a multi-relationship transaction topology graph with each transaction as a node and including semantic edges of role sharing and role exchange. The illegal resale detection module is used to input the multi-relationship transaction topology graph into the data illegal resale detection model based on the enhanced relational graph convolutional network, in order to detect illegal resale in the transaction information.

Images