Vehicle transaction risk intelligent early warning and credit assessment method based on big data analysis

Abstract

The invention discloses a vehicle transaction risk intelligent early warning and credit assessment method based on big data analysis. The method comprises the steps of S1, collecting an original transaction data set; s2, constructing a heterogeneous graph structure based on the original transaction data set; s3, dividing the heterogeneous graph structure into a graph snapshot sequence according to a fixed time window, and generating a time evolution graph sequence with time evolution characteristics; s4, constructing low-dimensional feature representations of the user nodes and the transaction event nodes under each time window; s5, constructing a causal path graph according to the time sequence and behavior dependency relationship of the vehicle transaction record data, and extracting a high-frequency transaction causal chain from the causal path graph; s6, performing joint modeling on the low-dimensional feature representation of the user node and the transaction event node and the high-frequency transaction causal chain; and S7, outputting the risk level label of the target transaction and the credit scoring result of the corresponding user. The method has the advantages of being high in structure expression ability, clear in causal logic and high in risk assessment accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent early warning, and in particular to an intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis. Background Art

[0002] With the rapid growth of the automobile ownership and the continuous expansion of the trading scale of the used car market, the vehicle transaction scenario faces a highly complex participant structure, diverse trading behavior characteristics and a market environment with highly asymmetric information. In the context of the lack of a unified credit assessment system and a dynamic risk identification mechanism, how to accurately model the behavior patterns of transaction participants, early warn of transaction risks, and scientifically evaluate user credit has become one of the core issues in the fields of intelligent transportation and intelligent transaction supervision.

[0003] In the prior art, vehicle transaction risk assessment mainly relies on static rule engines or expert experience modeling methods. The core means are mostly scoring systems that combine behavior feature extraction and threshold judgment, which have obvious deficiencies such as single evaluation dimension, slow update of risk features, and lack of interpretability of the scoring mechanism. Especially in the face of heterogeneous transaction graph structures with complex participant identities and dense cross-links in behavior chains, traditional models cannot effectively capture the dynamic correlation and evolution relationship between multi-source data, resulting in a decline in transaction risk identification ability, distortion of credit scoring results, and evaluation biases such as "high score but high risk" or "low score but stable reputation" are likely to occur. In addition, although some studies have tried to introduce machine learning and deep models for risk prediction, most of them only model at the feature level, lack the ability to model the evolution process of transaction structures, and do not integrate causal logic explanation mechanisms. They still stay in the "black box prediction" model stage and are difficult to support the actual needs of intelligent supervision scenarios for transaction traceability, risk interpretability, and decision intervention.

[0004] In terms of the decision support of risk scoring results, the prior art generally uses the method of dividing levels by fixed scoring intervals, which insufficiently considers the influence of different user behavior sequences, transaction paths and historical risk records, and lacks a flexible control mechanism that combines structural dynamic changes and risk propagation paths, resulting in poor real-time performance and differential expression ability of credit scoring, and is not conducive to realizing multi-dimensional risk perception and dynamic credit evolution control in complex transaction systems.

[0005] In summary, there is an urgent need for an intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis to achieve the deep integration of modeling the structural evolution characteristics, extracting risk causal paths and jointly evaluating risk credit in the vehicle transaction process, and effectively solve the core problems of static evaluation models, lagging risk identification and unexplainable scoring mechanisms existing in the prior art. Summary of the Invention

[0006] An object of the present invention is to propose an intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis. The present invention integrates a structurally adaptive heterogeneous graph neural network and a credit behavior causal reasoning engine, constructs a time-evolving graph modeling, node type-aware embedding, and high-frequency transaction causal chain extraction mechanism, and realizes the quantitative assessment of transaction risk levels and user credit scores through joint score graph generation and risk correlation function. It also realizes dynamic risk identification and interpretable credit modeling under complex transaction structures, and has the advantages of strong structural expression ability, clear causal logic, and high risk assessment accuracy.

[0007] A method for intelligent early warning and credit assessment of vehicle transaction risks based on big data analysis according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect vehicle transaction record data, user behavior trajectory data, third-party credit data, and social interaction data, and perform identification normalization and time series alignment to generate an original transaction data set;

[0009] S2. Based on the original transaction data set, construct a heterogeneous graph structure;

[0010] S3. Divide the heterogeneous graph structure into a sequence of graph snapshots according to a fixed time window, and perform structurally adaptive reconstruction on each graph snapshot in the sequence of graph snapshots based on the node addition rate and edge change rate to generate a time-evolving graph sequence with time-evolving characteristics;

[0011] S4. For each graph snapshot in the time-evolving graph sequence, use a graph neural network to perform node embedding calculation to generate low-dimensional feature representations of user nodes and transaction event nodes at each time window;

[0012] S5. Based on the original transaction data set and the time-evolving graph sequence, construct a causal path graph according to the chronological order and behavior dependence relationship of the vehicle transaction record data, and extract high-frequency transaction causal chains from it;

[0013] S6. Jointly model the low-dimensional feature representations of user nodes and transaction event nodes and the high-frequency transaction causal chains to generate a transaction score graph, and define a risk correlation function for calculating the vehicle transaction risk coefficient and user credit score value;

[0014] S7. Based on the transaction score graph, output the risk level label of the target transaction and the credit score result of the corresponding user.

[0015] Optionally, the original transaction data set includes vehicle transaction record data, user behavior track data, third-party credit data, and social interaction data, and performs identification normalization and time series alignment. Among them, the vehicle transaction record data is used to identify the transaction behavior relationship between users and vehicles, the user behavior track data is used to identify the account operation similarity relationship between different user accounts, and the social interaction data is used to mine the behavior co-occurrence relationship between users.

[0016] Optionally, the construction of the heterogeneous graph structure includes: constructing user nodes, vehicle nodes, transaction event nodes, and social nodes respectively with user entities, vehicle entities, transaction event entities, and social entities extracted from the original transaction data set. Based on the mapping relationship between user identifiers and vehicle identifiers included in the vehicle transaction record data, a transaction behavior relationship edge is generated between the user node and the vehicle node. Based on different user accounts with similar account operation characteristics in the user behavior track data, an account operation similarity relationship edge is constructed. Based on the message interaction behavior existing in the social interaction data and the user pairs jointly participating in the transaction behavior, a social co-occurrence relationship edge is generated, forming a heterogeneous graph structure containing multiple types of nodes and multiple types of relationship edges.

[0017] Optionally, S3 specifically includes:

[0018] S31. Time-slice the heterogeneous graph structure according to a fixed time window to obtain a graph snapshot sequence where G t represents the graph snapshot corresponding to the t-th time window, and n represents the total number of time windows;

[0019] S32. For each graph snapshot G t , calculate the node addition rate α t and the edge change rate β t :

[0020]

[0021] where V t and E t respectively represent the node set and the edge set of the graph snapshot G t , V t-1 and E t-1 respectively represent the node set and the edge set of the graph snapshot of the previous time window. The symbol \ represents the set difference operation, and the symbol △ represents the symmetric difference of the sets;

[0022] S33. When the following adaptive reconstruction condition is met:

[0023] α t > θ v ∪β t > θ e ;

[0024] Then, perform a structural update on the graph snapshot G t to update the adjacency matrix and node attribute representation of the graph snapshot G t , where θ v and θ e are the node addition rate threshold and edge change rate threshold respectively;

[0025] S34. Arrange all the graph snapshots after structural update in chronological order to form a time-evolution graph sequence where G' t represents the graph snapshot after structure adaptive reconstruction, m represents the number of updated graph snapshots, and m ≤ n.

[0026] Optionally, the specific steps of S33 include:

[0027] S331. For all nodes v t in the graph snapshot G i that meet the structure adaptive reconstruction conditions, perform structure perturbation detection, compare the current neighbor set with the neighbor set t-1 in the previous graph snapshot G to construct a perturbed neighbor set

[0028] S332. For each node v t in the graph snapshot G i , calculate the node representation after structure update based on the structure-aware weighted aggregation result of the current node representation and the perturbed neighbor node set . Its expression is:

[0029]

[0030] where γ ∈ [0, 1] is the structure perturbation fusion factor, represents the node representation of the perturbed neighbor node v j at time t, is the perturbed structure influence weight of node v j on node v i , and satisfies:

[0031] S333. Based on the node representation update result, reconstruct the adjacency matrix and edge set of the graph snapshot G t , retain the edge structure with a change intensity exceeding the edge perturbation threshold θ s to generate the graph snapshot G' t .

[0032] Optionally, the specific steps of S4 include:

[0033] S41. Initialize the embeddings of the user nodes and transaction event nodes in each graph snapshot G' after structure - adaptive reconstruction in the time - evolution graph sequence, and set the initial representation vectors according to the node types. t in, respectively, and set the initial representation vectors according to the node types.

[0034] S42. Before the start of each propagation layer l, based on the distribution structure of heterogeneous node types in the graph snapshot, execute the type - constrained sampling strategy of neighbor nodes to construct the type - aware neighbor set of each node in the propagation layer.

[0035] S43. For each node v i , construct the propagation input according to the type - aware neighbor set , collect the representations of neighbor nodes in the previous layer and retain the node type information;

[0036] S44. Use the node - type - aware aggregation mechanism to update the node representations:

[0037]

[0038] where, represents the representation vector of node v i in the l - th layer, represents the representation vector of node v j in the previous propagation layer l - 1, represents the set of neighbor nodes of node v i in the current graph snapshot, σ(·) represents the non - linear activation function, and MEAN(·) represents the aggregation function that performs element - wise averaging on all input vectors in the set. represents the trainable weight matrix corresponding to the node type φ(i) of node v i in the l - th layer, and φ(i) is the node - type mapping function;

[0039] S45. In each propagation layer l of performing graph neural network embedding calculation on the graph snapshot G', set the node representation dimension and activation function type of this propagation layer, and configure the corresponding type - propagation weight matrix in this propagation layer for different types of graph nodes to control the representation update method of each type of graph node in this propagation layer. t in this propagation layer for different types of graph nodes to control the representation update method of each type of graph node in this propagation layer.

[0040] S46. After all propagation layers are executed, extract the final representations of the user nodes and transaction event nodes as the low - dimensional semantic representations of the user nodes and transaction event nodes.

[0041] Optionally, the S42 specifically includes:

[0042] S421. For each graph snapshot G' after the structural adaptive reconstruction of each warp, t for each target graph node v i , obtain the type φ(i) of this node, and extract the set of neighbor nodes that have connection edges with this node in the graph as the candidate neighbor set;

[0043] S422. Construct a type constraint matrix between node types where K is the total number of node types in the graph, and Θ ab ∈[0,1] represents the weight coefficient that a graph node of type a allows sampling a graph node of type b;

[0044] S423. According to the type φ(i)=a of the target graph node v i , for the graph node v of type b in the candidate neighbor node set assign a sampling probability p j ; ij

[0045] S424. Perform a sampling operation on the candidate neighbor set ij according to the sampling probability p to obtain the type-aware neighbor set of the target graph node v i in the current propagation layer l The number of sampled nodes is controlled by the hyperparameter d;

[0046] S425. To improve the stability of the sampling process, perform multiple rounds of resampling and count the frequency of candidate nodes being sampled. If the frequency deviation exceeds the set threshold ∈, then renormalize the sampling probability distribution p of the candidate neighbors ij .

[0047] Optionally, the specific steps of S5 are as follows:

[0048] S51. Extract the transaction event nodes that occur in each time window from the time-evolving graph sequence , and combine the timestamps in the original transaction data set to construct a global event sequence {e1, e2,..., e n} in chronological order for all transaction events, where e i represents the i-th transaction event node, and n represents the total number of events;

[0049] S52. In each graph snapshot G' after the structural adaptive reconstruction of each warp t , based on the structural connection relationship and behavioral attribute dependence between the transaction event nodes in the graph structure, determine whether the event pairs in the global event sequence satisfy the causal trigger condition. If event e i precedes e j in time and in G't If there is a structural behavior interaction, then add a causal edge (e i →e j ) to the causal path graph;

[0050] S53. While adding the causal edge (e i →e j ), further assign a causal strength weight w ij , w ij ∈[0,1] to the causal edge:

[0051] w ij =α·freq(e i ,e j )+β·risk(e i );

[0052] Among them, freq(e i ,e j ) represents the historical co-occurrence frequency of the event pair (e i ,e j ) being identified as a causal edge in the time evolution graph sequence, risk(e i ) represents the risk value of the transaction behavior corresponding to the event e i , and α and β are weight balance coefficients.

[0053] S54. Calculate the cumulative value of all the weights on each causal edge with an assigned causal strength weight in the causal path graph, and set a causal strength threshold. Extract the causal edges with a cumulative weight value not lower than the causal strength threshold as the high-frequency trading causal chain.

[0054] Optionally, the specific steps of S6 include:

[0055] S61. Jointly model the final embedding representations of the user node and the transaction event node and the statistical feature vector c i composed of the causal edge frequency and edge weight information in the associated high-frequency trading causal chain to generate the joint feature representation z i of the graph nodes in the transaction scoring graph;

[0056] S62. Based on the joint feature representation z i , construct a transaction scoring graph and calculate the risk association score r i of each edge (v j ,v ij ) in the transaction scoring graph:

[0057]

[0058] Among them, r ij ∈[0,1] represents the degree of risk association between graph node pairs, zi With z j respectively represent the joint feature representations of graph nodes v i and v j of, M is a trainable scoring weight matrix, w ij is the causal edge weight between node pairs in the causal path graph, λ ∈ [0, 1] is the causal weight adjustment coefficient, and σ(·) is the Sigmoid function.

[0059] S63. Based on the risk association scores between user nodes and transaction event nodes in the scoring graph, calculate the vehicle transaction risk coefficient and the user credit score value respectively, and output the risk level labels of each transaction behavior and the credit score results of each user.

[0060] Optionally, the S63 specifically includes:

[0061] S631. Based on the risk association scores r ij between all user nodes and the target transaction event node in the transaction scoring graph, calculate the vehicle transaction risk coefficient R j of the transaction event node v j :

[0062]

[0063] Among them, represents the set of all user nodes that have scoring edge connections with the transaction event node v j , r ij represents the risk association score between the user node v i and the transaction event node v j , R j represents the risk coefficient of the transaction event, and the value range is from 0 to 1;

[0064] S632. Based on the risk association scores r ij between all transaction event nodes and the target user node in the transaction scoring graph, calculate the user credit score value C i of the user node v i :

[0065]

[0066] Among them, represents the set of all transaction event nodes that have scoring edge connections with the user node v i , C i represents the credit score value of the user, and the value range is from 0 to 1;

[0067] S633. According to the vehicle transaction risk coefficient R j and the user credit score value C iBased on the value of j , the corresponding risk level label is output for each transaction behavior node, and the corresponding credit score result is output for each user node. The risk level label is divided into three levels. When the risk coefficient R j is greater than 0.75, it is labeled as a high-risk level. When the risk coefficient R j is greater than 0.4 and less than or equal to 0.75, it is labeled as a medium-risk level. When the risk coefficient R

[0068] is less than or equal to 0.4, it is labeled as a low-risk level.

[0069] The beneficial effects of the present invention are as follows:

[0070] (1) In the process of heterogeneous graph modeling, the present invention introduces a structure adaptive reconstruction mechanism, constructs a graph snapshot structure perturbation detection and time-evolving graph sequence generation process based on the node addition rate and edge change rate, can dynamically reflect the structural change trend of vehicle transaction behaviors in the time dimension, effectively overcomes the limitation of the existing method's dependence on static graph structure modeling, improves the ability to capture the time-evolving characteristics of complex transaction behavior chains, and provides a structural basis for subsequent time-series risk identification and credit evolution assessment.

[0071] (2) By constructing a type-aware multi-layer graph neural network propagation mechanism, combining node initialization, neighbor type sampling and heterogeneous weight aggregation strategies, the present invention generates low-dimensional semantic embedding representations of user nodes and transaction event nodes, which have strong semantic distinguishability and high structural context consistency, breaks through the bottleneck of traditional embedding models with semantic ambiguity and poor representation consistency in heterogeneous graphs, and provides a high-expressive feature basis for risk association modeling and scoring among multi-type graph nodes. Description of the Drawings

[0072] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0073] Figure 1 is the overall flowchart of a vehicle transaction risk intelligent early warning and credit assessment method proposed by the present invention;

[0074] Figure 2Schematic diagram of the processing flow for constructing a time-evolving graph sequence based on a structure adaptive evolution mechanism for a method for intelligent early warning of vehicle transaction risks and credit assessment based on big data analysis proposed by the present invention;

[0075] Figure 3 Schematic diagram of the structure for generating risk labels and credit scores by jointly scoring a fusion type-aware graph neural network and a high-frequency trading causal chain for a method for intelligent early warning of vehicle transaction risks and credit assessment based on big data analysis proposed by the present invention. Detailed implementation manners

[0076] The present invention will now be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0077] Refer to Figures 1 - 3 , a method for intelligent early warning of vehicle transaction risks and credit assessment based on big data analysis, comprising the following steps:

[0078] S1. Collect vehicle transaction record data, user behavior trajectory data, third-party credit data, and social interaction data, and perform identity normalization and time series alignment to generate an original transaction data set;

[0079] S2. Based on the original transaction data set, construct a heterogeneous graph structure;

[0080] S3. Divide the heterogeneous graph structure into a graph snapshot sequence according to a fixed time window, and perform structure adaptive reconstruction on the graph snapshots in each graph snapshot sequence based on the node addition rate and edge change rate to generate a time-evolving graph sequence with time evolution characteristics;

[0081] S4. For each graph snapshot in the time-evolving graph sequence, use a graph neural network to perform node embedding calculation to generate low-dimensional feature representations of user nodes and transaction event nodes under each time window;

[0082] S5. Based on the original transaction data set and the time-evolving graph sequence, construct a causal path graph according to the time sequence and behavior dependency relationship of the vehicle transaction record data, and extract the high-frequency trading causal chain therefrom;

[0083] S6. Jointly model the low-dimensional feature representations of user nodes and transaction event nodes with the high-frequency trading causal chain to generate a transaction scoring graph, and define a risk correlation function to calculate the vehicle transaction risk coefficient and the user credit score value;

[0084] S7. Based on the transaction scoring graph, output the risk level label of the target transaction and the credit score result of the corresponding user.

[0085] Through the multi-source fusion and temporal alignment of vehicle transaction record data, user behavior trajectory data, third-party credit data, and social interaction data in the embodiments of the present invention, a unified abstract representation of the original transaction behavior can be achieved; combined with the structure-adaptive heterogeneous graph construction mechanism and the generation method of the time-evolving graph snapshot sequence, the dynamic behavior evolution pattern between transaction participants can be effectively captured; through the semantic separation low-dimensional embedding realized by the node type-aware graph neural network propagation process, the class distinguishability and structural stability of node feature expression are improved; on this basis, the causal path graph constructed based on the transaction record time sequence and structural interaction relationship, combined with the high-frequency causal edge extraction mechanism and the causal weight design strategy, forms an interpretable risk path representation; finally, through the transaction scoring graph generation process, the transaction risk and user credit characteristics are jointly modeled, and the vehicle transaction risk level and user credit score results are output, realizing a closed-loop linkage processing from risk prediction, causal modeling to scoring explanation, and significantly enhancing the dynamic expression ability, causal interpretability and practicality of joint evaluation of the model in a multi-source heterogeneous data environment.

[0086] In this embodiment, the original transaction data set includes collecting vehicle transaction record data, user behavior trajectory data, third-party credit data, and social interaction data, and performing identity normalization and temporal alignment. Among them, the vehicle transaction record data is used to identify the transaction behavior relationship between the user and the vehicle, the user behavior trajectory data is used to identify the account operation similarity relationship between different user accounts, and the social interaction data is used to mine the behavior co-occurrence relationship between users.

[0087] The present invention effectively realizes the multi-dimensional behavior network expression from transaction behavior, account operation mode to social behavior association by normalizing and fusing vehicle transaction record data, user behavior trajectory data, third-party credit data, and social interaction data; among them, the vehicle transaction record data is used to explicitly define the transaction behavior relationship between the user and the vehicle, the user behavior trajectory data is used to describe the operation similarity between different user accounts, and the social interaction data is used to discover the potential behavior co-occurrence relationship between users in the social chain. This original data fusion mechanism provides a high-quality, structurally clear multi-relationship basis for the subsequent heterogeneous graph structure construction, and significantly improves the system's ability to identify risk nodes and potential abnormal patterns in vehicle transaction behaviors.

[0088] In this embodiment, the construction of the heterogeneous graph structure includes: constructing user nodes, vehicle nodes, transaction event nodes, and social nodes respectively with the user entities, vehicle entities, transaction event entities, and social entities extracted from the original transaction data set, generating transaction behavior relationship edges between user nodes and vehicle nodes based on the mapping relationship between user identifiers and vehicle identifiers included in the vehicle transaction record data, constructing account operation similarity relationship edges between different user accounts with similar account operation characteristics in the user behavior trajectory data, and generating social co-occurrence relationship edges between user pairs with message interaction behaviors and co-participation in transaction behaviors existing in the social interaction data, so as to form a heterogeneous graph structure including multiple types of nodes and multiple types of relationship edges.

[0089] In the present invention, by introducing multiple types of graph node entities such as user nodes, vehicle nodes, transaction event nodes, and social nodes when constructing the heterogeneous graph structure, and based on the transaction mapping relationship between users and vehicles, the similarity relationship of account operation behaviors, and the social interaction co-occurrence relationship between users, transaction behavior relationship edges, account similarity relationship edges, and social co-occurrence relationship edges are respectively constructed, forming a heterogeneous graph model with multiple types of relationships with clear structure and semantics; this structure effectively supports the complex behavior modeling of vehicle transaction participants in different dimensions, enables the system to shift from static entity description to relationship-driven modeling, lays a multi-graph structure foundation for subsequent structure evolution modeling and graph embedding propagation, and significantly enhances the model's expression ability for complex transaction behavior graphs and the ability to retain structural context.

[0090] In this embodiment, the specific steps of S3 include:

[0091] S31. Perform time slicing on the heterogeneous graph structure according to a fixed time window to obtain a sequence of graph snapshots where G t represents the graph snapshot corresponding to the t-th time window, and n represents the total number of time windows;

[0092] S32. For each graph snapshot G t , calculate the node addition rate α t and the edge change rate β t :

[0093]

[0094] where V t and E t respectively represent the node set and edge set of the graph snapshot G t , V t-1 and E t-1 respectively represent the node set and edge set of the graph snapshot of the previous time window, the symbol \ represents the set difference operation, and the symbol △ represents the symmetric difference;

[0095] Through the formulas of the node addition rate and the edge change rate, the structural changes between each graph snapshot in the time-evolving graph sequence and the previous time window can be quantitatively expressed, enabling the system to sensitively perceive the node addition trend and the dynamic changes in edge connections at the graph structure level; based on the set difference and normalization ratio calculation methods, these formulas construct a measurement mechanism for the relative change amplitude of structural perturbations, which can avoid the absolute offset error caused by different total numbers of nodes or edge densities, effectively improving the stability and generality of the graph reconstruction judgment criteria, providing clear, controllable, and adjustable dynamic trigger criteria for subsequent execution of structure adaptive reconstruction and perturbation detection, and enhancing the scientificity and executability of structural change decisions in time-evolving modeling.

[0096] S33. When the following adaptive reconstruction conditions are met:

[0097] α t > θ v ∪ β t > θ e ;

[0098] Then perform a structure update on the graph snapshot G t Update the adjacency matrix and node attribute representation of the graph snapshot G t where θ v and θ e are the node addition rate threshold and the edge change rate threshold respectively;

[0099] S34. Arrange all the graph snapshots after structure update in chronological order to form a time-evolving graph sequence where G' t represents the graph snapshot after structure adaptive reconstruction, m represents the number of updated graph snapshots, and m ≤ n.

[0100] In this embodiment, S33 specifically includes:

[0101] S331. For all nodes v t in the graph snapshot G i that meet the adaptive reconstruction conditions, perform a structure perturbation detection, compare the current neighbor set with the neighbor set t-1 in the previous graph snapshot G to construct a perturbed neighbor set

[0102] S332. For each node v t in the graph snapshot G i , based on the structure-aware weighted aggregation result of the current node representation and the perturbed neighbor node set , calculate the node representation after structure update, and its expression is:

[0103]

[0104] Among them, γ ∈ [0, 1] is the structural perturbation fusion factor. represents the perturbed neighbor node v j 's node representation at time t. is the weight of the perturbed structural influence of node v j on node v i and satisfies:

[0105] Through the node update formula based on weighted aggregation of structurally perturbed neighbors, the dynamic adjustment mechanism of node representations in the case of structural changes in the graph snapshot can be refined and modeled; this formula fuses the representation of the current node itself and the representation of the set of perturbed neighbors, and controls the influence weights of the two through the structural perturbation fusion factor. At the same time, the normalized perturbed influence weights are used to ensure the stability and controllability of the aggregation process, thereby realizing the smooth evolution and information absorption of nodes after structural perturbation; this design effectively improves the coherence and sensitivity of node representation updates under graph structure changes, can introduce new neighbor semantics while maintaining the original structural information, provides a stable node-level update basis for modeling the structural continuity in the evolution process of graph snapshots, and enhances the expression robustness of graph neural networks in heterogeneous dynamic graph environments.

[0106] S333. Reconstruct the adjacency matrix and edge set of the graph snapshot G t based on the node representation update result, and retain the edge structure with a change intensity exceeding the edge perturbation threshold θ s to generate the structurally updated graph snapshot G'. t .

[0107] In the process of graph structure modeling of the present invention, a structural perturbation detection mechanism based on the node addition rate and edge change rate is introduced. The heterogeneous graph is divided into a sequence of graph snapshots by using the time slicing strategy, and a perturbation threshold determination and structural update operation are performed on each graph snapshot, realizing the dynamic adaptive reconstruction of the graph structure over time; further, by performing neighbor set change detection on the nodes of the graph snapshot that meets the reconstruction conditions, and performing structure-aware weighted aggregation based on the set of perturbed neighbors to generate the embedded representation of the nodes after structural update, and then reconstructing the adjacency matrix and edge set based on the node representation update result, finally generating a time-evolving graph sequence containing graph evolution features; this mechanism effectively improves the modeling ability of graph neural networks for the characteristics of vehicle trading behavior evolving over time, enhances the adaptability of node representations to graph structure changes, and provides a dynamically consistent graph expression basis for subsequent semantic embedding and causal path extraction.

[0108] In this embodiment, the specific content of S4 includes:

[0109] S41. Initialize the embeddings of the user nodes and transaction event nodes in each graph snapshot \(G'\) after structure - adaptive reconstruction in the time - evolution graph sequence, and set the initial representation vectors according to the node types. t in the user nodes and transaction event nodes are respectively initialized for embedding, and the initial representation vectors are set according to the node types.

[0110] S42. Before each propagation layer \(l\), based on the distribution structure of heterogeneous node types in the graph snapshot, execute the type - constrained sampling strategy for neighbor nodes to construct the type - aware neighbor set of each node in the propagation layer.

[0111] S43. For each node \(v\) i , construct the propagation input according to the type - aware neighbor set , collect the representations of neighbor nodes in the previous layer and retain the node type information.

[0112] S44. Use the node - type - aware aggregation mechanism to update the node representations:

[0113]

[0114] where, represents the representation vector of node \(v\) i in the \(l\) - th layer, represents the representation vector of node \(v\) j in the previous propagation layer \(l - 1\), represents the set of neighbor nodes of node \(v\) i in the current graph snapshot, \(\sigma(\cdot)\) represents the non - linear activation function, \(MEAN(\cdot)\) represents the aggregation function that performs element - wise averaging on all input vectors in the set, represents the trainable weight matrix corresponding to the node type \(\varphi(i)\) of node \(v\) i in the \(l\) - th layer, and \(\varphi(i)\) is the node - type mapping function;

[0115] Through the node representation update mechanism, by averaging and aggregating the features of neighbor nodes of the target node in the current graph snapshot and jointly updating them in combination with the historical features of the target node itself, and introducing a node - type - aware weight matrix, different types of nodes can use corresponding independent parameters for modeling during propagation, thus realizing the semantic difference modeling ability of node representations in heterogeneous graphs; this mechanism avoids the problem of semantic confusion under heterogeneous structures in the homogeneous aggregation method, improves the expression accuracy of node features, and enhances the generalization ability of the model in the distribution structure of multi - type nodes, providing a stable and distinguishable node embedding expression basis for subsequent scoring and risk assessment.

[0116] S45. In the graph snapshot \(G'\) tIn each propagation layer l that performs graph neural network embedding calculation, set the node representation dimension and activation function type of this propagation layer, and configure the corresponding type propagation weight matrix in this propagation layer for different types of graph nodes, which is used to control the representation update method of each type of graph node in this propagation layer.

[0117] S46. After all propagation layers are executed, respectively extract the final representations of user nodes and transaction event nodes as the low-dimensional semantic representations of user nodes and transaction event nodes.

[0118] In this embodiment, the S42 specifically includes:

[0119] S421. In each graph snapshot G' after structure adaptive reconstruction t , for each target graph node v i , obtain the type φ(i) of this node, and extract the set of neighbor nodes that have connection edges with this node in the graph as the candidate neighbor set;

[0120] S422. Construct a type constraint matrix between node types where K is the total number of node types in the graph, and Θ ab ∈[0,1] represents the weight coefficient that a graph node of type a allows to sample a graph node of type b;

[0121] S423. According to the type φ(i)=a of the target graph node v i , for the graph node v of type b in the candidate neighbor node set j allocate the sampling probability p ij ;

[0122] S424. Perform a sampling operation on the candidate neighbor set ij according to the sampling probability p to obtain the type-aware neighbor set i of the target graph node v in the current propagation layer l The number of sampled nodes is controlled by the hyperparameter d;

[0123] S425. To improve the stability of the sampling process, perform multiple rounds of resampling and count the frequencies of candidate nodes being sampled. If the frequency deviation exceeds the set threshold ∈, then renormalize the sampling probability distribution p ij .

[0124] In the process of generating the graph node embedding representation, the present invention introduces a type-aware graph neural network propagation mechanism, constructs a multi-layer embedding calculation process from node initialization, type-constrained sampling to weight aggregation and update, and can deeply model the behavioral structure context of user nodes and transaction event nodes in the heterogeneous graph; by constructing a type-aware neighbor set for each propagation layer, semantic constraints on the propagation adjacency range are realized; further, based on the type association matrix, the cross-type sampling probability is calculated to generate a high-fidelity structure context neighbor set, and sampling bias verification is performed in multiple rounds of sampling, enhancing the stability of adjacency modeling and the type separation ability; in addition, a node type-aware weight matrix is introduced in the update of node propagation expressions, and a type-based parameter sharing and non-sharing strategy is constructed to realize differential control of feature expressions for different node types; finally, low-dimensional semantic representations of user nodes and transaction event nodes are output, providing a basis for graph embedding expressions with stable structure and clear semantics for subsequent joint modeling, and improving the embedding expression ability and generalization performance of the system in a complex heterogeneous graph environment.

[0125] In this embodiment, the S5 specifically includes:

[0126] S51. Extract the transaction event nodes occurring in each time window from the time-evolving graph sequence , and combine the timestamps in the original transaction data set to construct a global event sequence {e1, e2,..., e n} in chronological order for all transaction events, where e i represents the i-th transaction event node and n represents the total number of events;

[0127] S52. In each graph snapshot G' t after structure adaptive reconstruction, based on the structural connection relationship and behavioral attribute dependence between transaction event nodes in the graph structure, judge whether the event pairs in the global event sequence satisfy the causal trigger condition. If event e i is earlier than e j in time and there is a structural behavior interaction in G' t , then add a causal edge (e i → e j ) to the causal path graph;

[0128] S53. When adding the causal edge (e i → e j ), further assign a causal strength weight w ij to the causal edge, where w ij ∈ [0, 1]:

[0129] w ij = α · freq(e i , e j ) + β · risk(ei );

[0130] where freq(e i , e j ) represents the historical co-occurrence frequency of the event pair (e i , e j ) being recognized as a causal edge in the time-evolution graph sequence, risk(e i ) represents the risk value of the trading behavior corresponding to event e i , and α and β are weight balance coefficients.

[0131] Through the causal edge strength weight assignment mechanism, the system can not only identify the causal relationships between trading events, but also quantitatively express the strength of each causal edge according to the historical frequency and risk impact degree. This weight mechanism integrates the co-occurrence frequency of event pairs being recognized as causal edges in the historical graph sequence and the risk values of the trading behaviors corresponding to the cause events. Through a two-factor weighted modeling method, it effectively distinguishes stable propagation paths from occasional interference paths, making the causal path graph more discriminative and interpretable; this mechanism provides a basis for subsequent high-frequency causal chain screening and scoring graph construction, optimizing the edge weight expression based on risk logic, and enhancing the structural expression ability and decision-making rationality of causal reasoning in the credit assessment scenario.

[0132] S54. Calculate the cumulative value of all weights on each causal edge in the causal path graph that assigns causal strength weights, set a causal strength threshold, and extract the causal edges whose cumulative weight values are not lower than the causal strength threshold as high-frequency trading causal chains.

[0133] The present invention extracts the trading event nodes that occur in chronological order in the time-evolution graph sequence, constructs a global time sequence of trading events, and combines the structural connection relationship and the behavioral attribute dependence relationship to design an event pair determination mechanism based on causal trigger conditions, constructs a causal path graph, and identifies the causal chains with stable propagation characteristics therein, enhancing the system's modeling ability for potential risk propagation paths in complex trading structures; further, when adding causal edges, a causal edge weight mechanism is introduced, and a causal edge weight function is constructed by combining the event co-occurrence frequency and the risk degree, realizing the quantitative expression of the strength of causal edges, and being able to more accurately distinguish the key trigger points and marginal paths in risk propagation; finally, high-frequency causal chains are screened by setting a threshold for the cumulative value of causal edge weights in the causal path graph, significantly enhancing the ability to identify high-risk trading causal patterns, and providing clear, stable, and quantifiable causal basis for scoring graph modeling and risk interpretation.

[0134] In this embodiment, the specific steps of S6 include:

[0135] S61. The final embedding representations of the user node and the trading event node The statistical feature vector c composed of the causal edge frequency and edge weight information in the associated high-frequency trading causal chain i is jointly modeled to generate the joint feature representation z of the graph nodes in the trading score graph i ;

[0136] S62. Based on the joint feature representation z i , construct a trading score graph and calculate the risk association score r of each edge (v i , v j ) in the trading score graph ij :

[0137]

[0138] where r ij ∈[0,1] represents the degree of risk association between graph node pairs, z i and z j respectively represent the joint feature representations of graph nodes v i and v j , M is a trainable scoring weight matrix, w ij is the causal edge weight between node pairs in the causal path graph, λ∈[0,1] is the causal weight adjustment coefficient, and σ(·) is the Sigmoid function

[0139] Through the risk association degree calculation mechanism, the system can comprehensively consider the joint feature expression of graph nodes and the edge weight information in the causal path graph when constructing the trading score graph, and establish a quantization model for the risk propagation intensity. This mechanism realizes the fusion modeling of structural information, semantic representation and causal logic by jointly encoding the low-dimensional semantic representation of nodes and historical causal weights, and introducing trainable scoring weight parameters and causal contribution adjustment factors, avoiding the deviation problem that risk scoring only depends on node representation or static relationship. This mechanism improves the expression consistency of edge connections in the scoring graph at the semantic and causal levels, effectively enhances the ability to identify high-risk trading behavior paths, and provides stable support for risk visualization modeling and credit quantization assessment in the vehicle trading scenario

[0140] S63. Based on the risk association scores between user nodes and trading event nodes in the scoring graph, calculate the vehicle trading risk coefficient and user credit score value respectively, and output the risk level labels of each trading behavior and the credit score results of each user

[0141] In this embodiment, S63 specifically includes:

[0142] S631. Based on the risk association scores r ij between all user nodes and the target trading event node in the trading score graph, calculate the vehicle trading risk coefficient R of the trading event node v j ​j :

[0143]

[0144] Among them, represents the set of all user nodes that have a scoring edge connection with the transaction event node v j , and r ij represents the risk correlation score between the user node v i and the transaction event node v j . R j represents the risk coefficient of the transaction event, and its value range is from 0 to 1;

[0145] Through the vehicle transaction risk coefficient calculation mechanism, the system can quantify and output the overall risk level of the transaction behavior based on the risk correlation scores between all user nodes and the target transaction event node in the transaction scoring graph. This mechanism constructs an event-oriented aggregated risk expression by aggregating all user scoring contributions connected to the target transaction node in terms of structure, breaking through the problem that the traditional single-point scoring model cannot reflect the risk consensus from the group perspective; in addition, this risk coefficient has a standardized interval expression form, which is convenient for comparative analysis and level division among multiple transaction tasks, providing a clear and consistent quantitative basis for subsequent risk label annotation and credit assessment logical reasoning.

[0146] S632. Calculate the user credit score value C ij of the user node v i based on the risk correlation score r i between all transaction event nodes and the target user node in the transaction scoring graph:

[0147]

[0148] Among them, represents the set of all transaction event nodes that have a scoring edge connection with the user node v i , and C i represents the user's credit score value, and its value range is from 0 to 1;

[0149] Through the calculation mechanism of the user credit score value, the system can reverse-derive the overall credit level of the user in multiple transaction scenarios based on the risk correlation scores between the user node and all its associated transaction event nodes in the transaction scoring graph. This mechanism uses the mean of all transaction risk scores for reflection and performs standardization processing on the numerical expression, ensuring that the scoring results are comparable and interpretable on a unified scale; through the integration of structural scoring and the breadth of transaction behavior, this scoring mechanism effectively reflects the historical behavior risk contribution of the user, improves the evaluation ability of the model for the user's comprehensive credit performance, and provides a quantifiable support means for credit grading, behavior tracking, and risk control linkage mechanisms.

[0150] S633. According to the vehicle transaction risk coefficient R j and the user credit score value C i values, output corresponding risk level labels for each transaction behavior node, and output corresponding credit score results for each user node. The risk level labels are divided into three levels. When the risk coefficient R j is greater than 0.75, it is labeled as a high-risk level. When the risk coefficient R j is greater than 0.4 and less than or equal to 0.75, it is labeled as a medium-risk level. When the risk coefficient R j is less than or equal to 0.4, it is labeled as a low-risk level.

[0151] In the process of building the transaction scoring graph model of the present invention, the low-dimensional representations of user nodes and transaction event nodes are jointly encoded with the frequency and risk information of causal edges in the high-frequency transaction causal chain to generate a joint feature representation of graph nodes, improving the modeling accuracy of the coupling relationship between user behavior and transaction risk; further, by introducing a scoring weight matrix and a causal edge weight fusion factor, a risk correlation function in the transaction scoring graph is constructed to measure the risk propagation intensity between any node pairs in the graph, enhancing the semantic distinguishability of the edge connections in the scoring graph; based on the calculation of the scoring results, the vehicle transaction risk coefficient is calculated for the transaction event nodes respectively, and the user credit score value is calculated for the user nodes, and a multi-level risk level classification mechanism is designed to automatically label the transaction level according to the risk value threshold, and output structured credit evaluations and risk labels, breaking through the technical bottleneck that the traditional scoring system cannot simultaneously reflect the risk source propagation path and the interpretability of the scoring level classification, and improving the comprehensive performance and operability of multi-objective decision-making evaluation in the vehicle transaction field.

[0152] Example:

[0153] To verify the feasibility and actual effect of the present invention, the present invention is applied to the B city headquarters system of a large domestic used car trading platform. The average annual transaction volume of this platform exceeds 1.2 million times, and the number of daily active users is about 180,000, covering multiple business modules such as vehicle transactions, financial installments, insurance services, and car dealer ratings. Due to the high involved amount, complex participating entities, and frequent cross-platform collaborations, this platform has long faced security risks such as insufficient user credit assessment, difficulty in identifying suspicious transaction behaviors, and frequent vehicle transaction frauds. Especially in the case of imperfect user real-name authentication systems, there are a large number of abnormal behaviors such as manipulating transaction scores using multiple accounts, repeatedly posting false vehicle information, and maliciously creating transaction chains, and the traditional risk control system is difficult to trace the behavior chain and risk traceability path.

[0154] The platform decides to deploy the "Intelligent Early Warning and Credit Assessment Method for Vehicle Transaction Risks Based on Self-Evolving Heterogeneous Graph Neural Network and Credit Behavior Causal Inference Engine" proposed by the present invention in its risk control architecture to enhance the behavior modeling ability for multi-source heterogeneous data and the accuracy of risk identification.

[0155] During the deployment process, the system first accesses historical data for the past 12 months from the platform's transaction database, including vehicle transaction records, user behavior logs, third-party credit scores, and social interaction information inside and outside the platform. The system performs unified identification, time alignment, and feature normalization processing on these raw data, constructs a set of original transaction data, and based on this, establishes a heterogeneous graph structure, which includes user nodes, vehicle nodes, transaction event nodes, and social nodes. At the same time, multi-type edges are established based on relationships such as account similarity, social co-occurrence, and transaction behavior association to form a heterogeneous graph relationship network.

[0156] To dynamically model the evolution process of transaction behavior, the system slices the heterogeneous graph with a week as the time window to form a sequence of graph snapshots for each week. Based on the node addition rate and edge change rate in each graph snapshot, the system performs structural perturbation detection and graph structure adaptive reconstruction, automatically reconstructing the adjacency matrix and edge set within the window where significant structural changes are detected to form a time-evolving graph sequence with dynamic structural evolution characteristics. After that, based on the graph snapshots with updated structures, the system uses graph neural networks for node embedding propagation and adopts a type-aware multi-layer aggregation mechanism to generate low-dimensional semantic representations of user and transaction event nodes.

[0157] The system further combines the transaction time sequence and behavior causal dependence to identify the causal trigger paths between transaction events from the evolving graph and constructs a causal path graph. For the stable path chains that frequently appear among them, the system assigns causal weights jointly calculated based on frequency and risk, thereby screening out causal path subgraphs with high risk control value.

[0158] Finally, the system jointly models the semantic representations of user nodes and transaction nodes with the causal subgraph to generate a transaction scoring graph, calculates the risk association scores between each pair of nodes through a scoring function, further calculates the risk coefficient of each transaction behavior and the credit score of each user, and the system automatically outputs the corresponding risk level labels and user rating results.

[0159] After implementing this system, the platform conducted a three-month actual evaluation, mainly testing from three dimensions: risk identification ability, user credit score stability, and transaction fraud rate. The following is a comparison of the platform's key risk control indicators before and after deployment, as shown in Table 1:

[0160] Table 1: Comparison Table of Risk Control Performance Indicators for the Second-Hand Car Platform in City B

[0161]

[0162]

[0163] As can be seen from the above table, the deployment of the present invention has significantly improved the platform's ability to identify abnormal transactions. The system can identify 129 high-risk trading behaviors per month on average, which is about 2.7 times that before deployment. At the same time, due to the causal path explanation mechanism supporting more detailed risk cause analysis, the system issued automatic warnings for 312 abnormal trading behaviors within three months, assisting the risk control team to intervene in advance, and finally increasing the coverage rate of fraud behaviors intercepted and confirmed from 41.2% to 87.6%.

[0164] In addition, the stability of the user credit scoring mechanism has been significantly improved, and the standard deviation of the score has decreased from 0.61 to 0.23, indicating that the scoring system is more coherent and stable and is no longer easily affected by trading frequency and short-term behaviors and fluctuates violently. In terms of abnormal account identification, the system uses multi-source information graph structure joint modeling to identify the operational similarity and behavior co-occurrence relationship between multiple accounts, and increases the accuracy of account clustering identification from 65.7% to 93.2%, effectively combating some malicious behaviors in the platform that manipulate trading scores with sock puppet accounts.

[0165] After deployment, the overall response efficiency of the system is excellent, with an average time consumption of 4.6 seconds in the complete process of comprehensive graph calculation, embedding propagation, and risk reasoning, meeting the platform's requirements for real-time response. At the same time, after high-risk transactions are automatically marked by the system, the manual confirmation response time is shortened from the original 42 minutes to 6 minutes, greatly alleviating the workload of manual risk control on the platform.

[0166] In a typical case, the system found that a user posted transaction records of the same vehicle three times through different accounts within two weeks. Although the transaction amounts changed slightly, the operation time, geographical location, and social account correlation degree were highly consistent. Through comparison with the causal path graph, the system determined it as a typical abnormal trading behavior and issued a medium-risk warning in advance. After verification by the platform's risk control team, it was confirmed that the vehicle was an illegally mortgaged vehicle, successfully preventing a loan transaction of 50,000 yuan.

[0167] In summary, the vehicle trading risk assessment method based on self-evolving heterogeneous graph neural network and causal reasoning engine proposed by the present invention can automatically identify structural changes from multi-dimensional data, mine hidden causal chains, and provide interpretable risk warning results, significantly improving the platform's risk control ability, user identification accuracy, and response efficiency in actual business, verifying the effectiveness and practical value of the present invention.

[0168] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention should cover within the protection scope of the present invention according to the technical solution and inventive concept of the present invention for equivalent substitution or change.

Claims

1. An intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis, characterized in that, It includes the following steps: S1. Collect vehicle transaction record data, user behavior trajectory data, third-party credit data, and social interaction data, and perform identification normalization and time series alignment to generate an original transaction data set; S2. Based on the original transaction data set, construct a heterogeneous graph structure; S3. Divide the heterogeneous graph structure into a graph snapshot sequence according to a fixed time window, and perform structure adaptive reconstruction on the graph snapshots in each graph snapshot sequence based on the node addition rate and edge change rate to generate a time-evolution graph sequence with time evolution characteristics; S4. For each graph snapshot in the time-evolution graph sequence, use a graph neural network to perform node embedding calculation to generate low-dimensional feature representations of user nodes and transaction event nodes under each time window; S5. Based on the original transaction data set and the time-evolution graph sequence, construct a causal path graph according to the time sequence and behavior dependence relationship of the vehicle transaction record data, and extract high-frequency transaction causal chains therefrom; S6. Jointly model the low-dimensional feature representations of user nodes and transaction event nodes and the high-frequency transaction causal chains to generate a transaction scoring graph, and define a risk correlation function to calculate the vehicle transaction risk coefficient and the user credit score value; S7. Based on the transaction scoring graph, output the risk level label of the target transaction and the credit score result of the corresponding user.

2. The intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 1, wherein The original transaction data set includes collecting vehicle transaction record data, user behavior trajectory data, third-party credit data, and social interaction data, and performing identification normalization and time series alignment. Among them, the vehicle transaction record data is used to identify the transaction behavior relationship between users and vehicles, the user behavior trajectory data is used to identify the account operation similarity relationship between different user accounts, and the social interaction data is used to mine the behavior co-occurrence relationship between users.

3. An intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 1, characterized in that The construction of the heterogeneous graph structure includes: respectively constructing user nodes, vehicle nodes, transaction event nodes, and social nodes with user entities, vehicle entities, transaction event entities, and social entities extracted from the original transaction data set, generating transaction behavior relationship edges between user nodes and vehicle nodes based on the mapping relationship between user identifiers and vehicle identifiers included in the vehicle transaction record data, constructing account operation similarity relationship edges between different user accounts with similar account operation characteristics in the user behavior trajectory data, and generating social co-occurrence relationship edges between user pairs with message interaction behavior and common participation in transaction behavior in the social interaction data to form a heterogeneous graph structure including multiple types of nodes and multiple types of relationship edges.

4. An intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 1, characterized in that The specific content of S3 includes: S31. Perform time slicing on the heterogeneous graph structure according to a fixed time window to obtain a sequence of graph snapshots where G t represents the graph snapshot corresponding to the t-th time window, and n represents the total number of time windows; S32. For each graph snapshot G t , calculate the node addition rate α t and the edge change rate β t : Among them, V t and E t respectively represent the node set and edge set of the graph snapshot G t , V t-1 and E t-1 respectively represent the node set and edge set of the graph snapshot in the previous time window. The symbol \ represents the set difference operation, and the symbol △ represents the symmetric difference of sets; S33. When the following adaptive reconstruction conditions are met: α t >θ v ∪β t >θ e ; Then, for the graph snapshot G t perform a structural update to update the adjacency matrix and node attribute representation of the graph snapshot G t , where θ v and θ e are the node addition rate threshold and the edge change rate threshold, respectively; S34. Arrange all the graph snapshots with updated structures in chronological order to form a time-evolution graph sequence where G' t represents the graph snapshot after structure adaptive reconstruction, and m represents the number of updated graph snapshots, satisfying m ≤ n.

5. An intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 1, characterized in that, The specific content of S33 includes: S331. For all nodes v in the graph snapshot G that meet the adaptive reconstruction conditions t perform structural perturbation detection, and compare the current neighbor set i with the neighbor set in the previous graph snapshot G to construct a perturbed neighbor set t-1 ​​ S332. For the graph snapshot G t for each node v i , based on the current node representation and the structure-aware weighted aggregation result of the perturbed neighbor node set , calculate the node representation after structure update whose expression is: Among them, γ∈[0,1] is the structure perturbation fusion factor, represents the perturbed neighbor node v j at the node representation at time t, is the node v j for the node v i the perturbation structure influence weight of, satisfying: S333. Reconstruct the graph snapshot G based on the updated result of the node representation t of the adjacency matrix and edge set, and retain the edge structure with a change intensity exceeding the edge perturbation threshold θ s to generate the graph snapshot G' after structure update t .

6. The intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 1, characterized in that, The specific content of S4 includes: S41. Initialize the embeddings of the user nodes and transaction event nodes in each graph snapshot G' after structure adaptive reconstruction in the time evolution graph sequence, and set the initial representation vectors according to the node types t in it S42. Before the start of each propagation layer l, based on the distribution structure of heterogeneous node types in the graph snapshot, execute the type-constrained sampling strategy for neighbor nodes to construct the type-aware neighbor set of each node in the propagation layer S43. For each node v i , construct a propagation input based on the type-aware neighbor set , collect the representations of neighbor nodes in the previous layer and retain the node type information; S44. Use a node type-aware aggregation mechanism to update the node representation: Among them, represents the representation vector of node v in the l-th layer i and represents the representation vector of node v j in the previous propagation layer l - 1. represents the set of neighbor nodes of node v i in the current graph snapshot. σ(·) represents the non-linear activation function, and MEAN(·) represents the aggregation function that performs element-wise averaging operations on all input vectors in the set. represents the trainable weight matrix corresponding to the node type φ(i) to which node v i belongs in the l-th layer, where φ(i) is the node type mapping function; S45. In each propagation layer l that performs graph neural network embedding calculation on the graph snapshot G' t Set the node representation dimension and activation function type of the propagation layer, and configure the corresponding type propagation weight matrix in the propagation layer for different types of graph nodes to control the representation update method of each type of graph node in the propagation layer. S46. After the execution of all the transport layers is completed, the final representations of the user nodes and the transaction event nodes are respectively extracted as the low-dimensional semantic representations of the user nodes and the transaction event nodes.

7. An intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 6, characterized in that, The specific content of S42 includes: S421. In each graph snapshot G' after the adaptive reconstruction of the warp structure t for each target graph node v i obtain the type φ(i) of the node, and extract the set of neighbor nodes that have connection edges with the node in the graph as the candidate neighbor set; S422. Construct a type constraint matrix between node types where K is the total number of node types in the graph, and Θ ab ∈ [0, 1] represents the weight coefficient that a graph node of type a allows sampling a graph node of type b; S423. According to the type φ(i) = a of the target graph node v i , assign the sampling probability p to the graph node v j of type b in the candidate neighbor node set ij ; S424. According to the sampling probability p ij Sample the candidate neighbor set to obtain the target graph node v i in the type-aware neighbor set in the current propagation layer l The number of sampled nodes is controlled by the hyperparameter d; S425. To improve the stability of the sampling process, multiple rounds of resampling are performed and the sampling frequencies of candidate nodes are counted. If the frequency deviation exceeds the set threshold ∈, the sampling probability distribution p of candidate neighbors is renormalized ij .

8. A method for intelligent early warning of vehicle transaction risks and credit assessment based on big data analysis according to claim 1, characterized in that, The specific content of S5 includes: S51. Extract the transaction event nodes that occur within each time window from the time evolution graph sequence and, in combination with the timestamps in the original transaction data set, construct a global event sequence {e1, e2, …, e n} in chronological order for all transaction events, where e i represents the i-th transaction event node and n represents the total number of events; S52. In each graph snapshot G' after the structural adaptive reconstruction of each warp structure t , based on the structural connection relationship and behavioral attribute dependence between transaction event nodes in the graph structure, determine whether the event pairs in the global event sequence satisfy the causal trigger condition. If event e i is temporally prior to e j and there is a structural behavioral interaction in G' t , then add a causal edge (e i →e j ) to the causal path graph; S53. While adding a causal edge (e i →e j ), further assign a causal strength weight w ij to the causal edge, where w ij ∈ [0, 1]: w ij = α·freq(e i ,e j ) + β·risk(e i ); Among them, freq(e i , e j ) represents the historical co-occurrence frequency of the event pair (e i , e j ) being recognized as a causal edge in the time-evolution graph sequence, and risk(e i ) represents the risk value of the trading behavior corresponding to the event e i . α and β are weight balance coefficients. S54. Calculate the cumulative value of all weights on each causal edge with an assigned causal intensity weight in the causal path graph, and set a causal intensity threshold, and extract the causal edges with a weight cumulative value not lower than the causal intensity threshold as high-frequency transaction causal chains.

9. An intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 1, characterized in that, The specific content of S6 includes: S61. Jointly model the final embedding representations of the user node and the transaction event node with the statistical feature vector c composed of the causal edge frequencies and edge weight information in the associated high-frequency trading causal chain i to generate the joint feature representation z of the graph nodes in the transaction scoring graph i ; S62. Based on the joint feature representation z i , construct a transaction scoring graph and calculate the risk association score r i , v j ) of each edge (v ij ): Among them, r ij ∈ [0, 1] represents the degree of risk association between graph node pairs, z i and z j respectively represent the joint feature representations of graph nodes v i and v j , M is a trainable scoring weight matrix, w ij is the causal edge weight between node pairs in the causal path graph, λ ∈ [0, 1] is the causal weight adjustment coefficient, and σ(·) is the Sigmoid function. S63. Calculate the vehicle transaction risk coefficient and the user credit score value respectively based on the risk correlation score between the user node and the transaction event node in the scoring graph, and output the risk level label of each transaction behavior and the credit score result of each user.

10. The intelligent early warning and credit assessment method for vehicle transaction risks based on big data analysis according to claim 9, characterized in that, The specific content of S63 includes: S631. Based on the risk correlation score r between all user nodes and the target transaction event node in the transaction scoring graph ij , calculate the vehicle transaction risk coefficient R of the transaction event node v j : j ​ Among them, represents all user node sets that have a scoring edge connection with the transaction event node v j , r ij represents the user node v i 's risk association score with the transaction event node v j , R j represents the risk coefficient of the transaction event, and its value range is from 0 to 1; S632. Calculate the user credit score value C of user node v based on the risk correlation score r between all transaction event nodes and the target user node in the transaction score graph ij , calculate user node v i 's user credit score value C i : Among them, represents the set of all transaction event nodes that have a scoring edge connection with the user node v i C represents the set of all transaction event nodes that have a scoring edge connection with the user node v i represents the credit score value of the user, and the value range is from 0 to 1; S633. According to the vehicle transaction risk coefficient R j and the user credit score value C i values, output the corresponding risk level label for each transaction behavior node, and output the corresponding credit score result for each user node. The risk level label is divided into three levels. When the risk coefficient R j is greater than 0.75, it is marked as a high risk level. When the risk coefficient R j is greater than 0.4 and less than or equal to 0.75, it is marked as a medium risk level. When the risk coefficient R j is less than or equal to 0.4, it is marked as a low risk level.

Images