Analysis method for analyzing group associated transaction risk through machine learning

Abstract

The invention relates to the technical field of enterprise risk management, and particularly discloses an analysis method for analyzing group associated transaction risks through machine learning. According to the method, through enterprise group associated transaction multi-source heterogeneous data fusion enhancement, barrier breaking and information mining, comprehensive data support is provided, then a space-time diagram neural network and Transform mixed model is constructed, space-time features are extracted and fused, and comprehensive risk features are output; through weighted cross entropy in combination with contrast learning loss function training and AdamW optimizer parameter adjustment, decision interpretability is realized by means of SHAP and LIME; performing real-time evaluation and early warning, stream-oriented data calculation and processing, reinforcement learning threshold adjustment, and threshold-exceeding multi-channel push early warning; then, conducting prediction and intelligent decision making are carried out, Monte Carlo simulation risk conducting is carried out, and a deep reinforcement learning model outputs a prevention and control strategy; and finally, the model is continuously evolved, data drift is detected and updated, new data is adapted through increment and transfer learning, regular evaluation and optimization are carried out, and stable operation and sustainable development of an enterprise are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of enterprise risk management technology, specifically to an analytical method for analyzing the risks of related-party transactions within a group through machine learning. Background Technology

[0002] In the operation of corporate groups, related-party transactions, as an important means of internal resource allocation, can optimize resource allocation but may also bring significant risks due to acts such as improper transfer of benefits and financial fraud. Currently, the technical means for analyzing the risks of related-party transactions within a group have many limitations:

[0003] 1. Data Processing Level: Related-party transaction risk analysis involves multi-source data, including both structured and unstructured data. Traditional methods have weak capabilities in processing unstructured data (such as news, social media comments, and regulatory announcements), failing to effectively extract risk clues hidden in the text. Furthermore, the fusion of multi-source data suffers from inconsistent formats and semantic ambiguity, resulting in the data's value not being fully realized.

[0004] 2. Model Construction Level: Existing machine learning models mostly employ single structures or traditional algorithms, making it difficult to comprehensively capture the topological characteristics and time-series dynamic changes of interconnected transaction networks. For example, ordinary neural networks cannot effectively handle the complex relationship structures between nodes in transaction networks, and traditional time-series analysis methods are also unable to cope with the non-stationarity and suddenness of transaction data, resulting in low accuracy in risk identification.

[0005] 3. Interpretability: Most deep learning models are "black box" models, making it difficult to clearly explain the basis for risk assessment. Enterprise managers and regulatory agencies struggle to understand the model's decision-making logic, reducing the credibility and practicality of risk analysis results and hindering the formulation and implementation of risk prevention and control measures.

[0006] 4. Dynamic Adaptability: Business transaction models are constantly evolving, and new forms of risk are emerging one after another. Existing methods lack a dynamic update mechanism, cannot respond to changes in transaction models in a timely manner, and risk warnings are delayed, making it difficult to ensure the stable operation of enterprises.

[0007] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0008] The purpose of this invention is to provide an analytical method for analyzing the risks of related-party transactions within a group using machine learning, in order to solve the problems mentioned in the background above.

[0009] The objective of this invention can be achieved through the following technical solution: an analytical method for analyzing the risk of related-party transactions within a group using machine learning, comprising the following steps:

[0010] S1. Multi-source heterogeneous data fusion and enhancement: Comprehensive collection of related transaction data from within and outside the enterprise group; preprocessing of structured data using feature engineering techniques; semantic understanding and information extraction of unstructured data using pre-trained language models; data fusion by constructing a related transaction knowledge graph using graph databases; and enhancement processing of the fused data using generative adversarial networks and adversarial training techniques.

[0011] S2. Spatiotemporal Graph Neural Network Construction and Feature Extraction: A dynamic graph network is constructed with enterprises as nodes and related transaction relationships as edges. A spatiotemporal graph is formed according to the time window. A spatiotemporal graph neural network is designed. The local structural features of the nodes are learned through the graph attention mechanism of the spatial convolutional layer. The temporal convolutional layer uses the temporal convolutional network to extract time series features. The feature fusion layer fuses the two features and outputs a comprehensive risk feature representation of the nodes.

[0012] S3. Interpretable Hybrid Deep Learning Model Training: Construct a hybrid deep learning model that integrates spatiotemporal graph neural networks and Transformers. Optimize the loss using a weighted cross-entropy loss function combined with a contrastive learning loss function. Train the model using the AdamW optimizer and enhance model interpretability by combining SHAP values ​​and LIME techniques.

[0013] S4. Real-time risk assessment and dynamic early warning: Input the real-time collected related transaction data into the trained model, use the streaming computing framework to achieve real-time processing and inference, and dynamically adjust the risk assessment threshold based on historical data and market environment using reinforcement learning algorithms. When the risk probability exceeds the threshold, risk early warning information is automatically generated and pushed through multiple channels.

[0014] S5. Risk Transmission Prediction and Intelligent Decision-Making: Based on the trained model and spatiotemporal map, the Monte Carlo simulation method is used to predict the risk transmission path and impact range. An intelligent decision-making model based on deep reinforcement learning is constructed to automatically generate the optimal risk prevention and control strategy with the goal of minimizing risk loss and prevention and control costs.

[0015] S6. Continuous Model Evolution and Adaptation: A concept drift detection method based on dynamic time warping and cluster analysis is used to monitor data changes. Incremental learning and transfer learning techniques are used to update the model, and the model is evaluated and optimized regularly.

[0016] Optionally, the preprocessing of structured data using feature engineering techniques includes data cleaning, standardization, and feature encoding.

[0017] The unstructured data processing extracts key transaction information from the text using natural language processing (NLP) technology and transforms it into structured feature vectors. NLP technology includes, but is not limited to, named entity recognition (RANK) technology and relation extraction technology. Key transaction information includes, but is not limited to, transaction subject, transaction amount, transaction time, and risk keywords.

[0018] The construction of the related transaction knowledge graph involves using enterprises as nodes and related transactions as edges, constructing the related transaction knowledge graph using a graph database, assigning attributes to nodes and edges, where related transaction relationships include but are not limited to transaction relationships and equity relationships, assigning attributes to nodes including but not limited to enterprise name, industry type, and credit score attributes, and assigning attributes to edges including but not limited to transaction amount, transaction time, and pricing method attributes.

[0019] Optionally, the time window is divided into days, weeks, and months;

[0020] The spatial convolutional layer adaptively allocates attention weights based on the correlation strength and transaction characteristics between nodes;

[0021] The temporal convolutional layer captures the dynamic changes in trading risk over time, including the periodic fluctuations in trading frequency and the evolution patterns of risk events.

[0022] Optionally, the weighted cross-entropy loss function sets different weights for different categories according to the severity of the transaction risk;

[0023] The SHAP value quantifies the contribution of each feature to the risk prediction result from a global perspective;

[0024] The LIME generates interpretable linear models within a local scope.

[0025] Optionally, the risk warning information includes basic information about the risky transaction, risk level, explanation of risk characteristics, and the possible scope of risk impact, and displays the location of the risky transaction in the related transaction network and the risk propagation path on a visual interface.

[0026] Optionally, the Monte Carlo simulation method calculates the probability and potential loss of different nodes being affected by the risk at different time points by repeatedly simulating the risk diffusion process. The intelligent decision-making model takes the risk transmission prediction results as state input and the prevention and control measures as actions, and generates the optimal risk prevention and control strategy with the goal of minimizing risk loss and minimizing prevention and control costs.

[0027] Optionally, the concept drift detection method monitors the distribution changes and pattern evolution of transaction data in real time, and triggers a model update mechanism when a significant change in data distribution is detected; the incremental learning technology gradually integrates new data into the model training process, and the transfer learning technology transfers learned knowledge and features in similar transaction scenarios or industry data.

[0028] Optionally, the execution process of the triggering model update mechanism is as follows:

[0029] Obtain historical transaction dataset A hist Real-time transaction flow A real-time And divided into D according to the time window. t D t+1 ..., the historical transaction dataset is split according to transaction type, and the time-series feature sequences of each type are extracted. For real-time window D t Extracting time series sequences of the same type Calculated using DTW and The normalization distance is calculated and compared with a preset normalization distance threshold. If the normalization distance is greater than the preset normalization distance threshold, the transaction type is determined to be time-series drift.

[0030] For A hist Extract transaction features and use the DBSCAN algorithm to cluster them into K clusters. And record the density of each cluster. center For D t Using the same features and clustering parameters, clusters are obtained. And record the density of each cluster. center Calculate density change Center offset It is then compared with preset density change thresholds and center offset thresholds. If the number of clusters exceeding the thresholds is greater than K / 2, then feature distribution drift is determined.

[0031] If either time-series drift or feature drift is triggered, and the proportion of drift-covered transactions exceeds a preset threshold, then the model update mechanism is triggered to mark the drift type.

[0032] The beneficial effects of this invention are:

[0033] This invention breaks down data barriers through multi-source heterogeneous data fusion and enhancement technology, fully mining risk information from both structured and unstructured data. This improves data utilization to a certain extent, providing comprehensive and accurate data support for risk analysis, and making risk identification more extensive and in-depth. By employing a hybrid deep learning model combining spatiotemporal graph neural networks and Transformers, it effectively overcomes the limitations of traditional models, comprehensively capturing the topological characteristics and temporal dynamic changes of transaction networks, thus improving the accuracy and reliability of risk identification and risk warning. Furthermore, the introduction of SHAP and LIME technologies makes the model's decision-making process transparent, allowing enterprise managers and regulatory agencies to clearly understand the basis for risk assessment, thereby enhancing risk assessment capabilities. The reliability and practicality of the analysis results provide strong support for the formulation and implementation of risk prevention and control measures, and also help meet regulatory compliance requirements. Based on the dynamic threshold adjustment and continuous model evolution mechanism of reinforcement learning, the system can quickly adapt to changes in transaction patterns and new risks, and can promptly identify potential risks, giving enterprises more time to respond to risks and effectively improving their risk prevention and control capabilities and risk resistance levels. Through risk transmission prediction and intelligent decision-making models based on deep reinforcement learning, the system automates and intelligentizes risk prevention and control decisions, automatically generates optimal prevention and control strategies, reduces human decision-making errors, and to a certain extent reduces the losses caused by related-party transaction risks, improves the efficiency and effectiveness of enterprise risk management, and ensures the sound operation and sustainable development of enterprises. Attached Figure Description

[0034] The invention will now be further described with reference to the accompanying drawings.

[0035] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figure 1 As shown, this invention is an analytical method for analyzing the risk of related-party transactions within a group using machine learning, comprising the following steps:

[0038] S1. Multi-source heterogeneous data fusion and enhancement: Comprehensive collection of related transaction data from within and outside the enterprise group; preprocessing of structured data using feature engineering techniques; semantic understanding and information extraction of unstructured data using pre-trained language models; data fusion by constructing a related transaction knowledge graph using graph databases; and enhancement processing of the fused data using generative adversarial networks and adversarial training techniques.

[0039] It should be noted that, through standardized API interfaces, transaction records are collected from the enterprise group's internal financial system, including information such as transaction date, transaction amount, and accounts of both parties; contract texts are obtained from the contract management system, covering contract subject matter, transaction terms, liability for breach of contract, etc.; and logistics data, such as structured data like cargo transportation routes, warehousing information, and delivery time, are extracted from the supply chain management system.

[0040] Using web crawler technology, we scrape unstructured text data related to enterprises, such as news information and user comments, from news websites and social media platforms; through official data interfaces, we access regulatory announcements to obtain information such as enterprise violation records and regulatory penalties; at the same time, we connect with third-party credit rating agencies and business registration platforms to obtain external data such as enterprise credit scores, equity structure, legal representative information, and historical change records.

[0041] Specifically, the structured data is preprocessed using feature engineering techniques, including data cleaning algorithms to identify and remove duplicate transaction records. For missing values, imputation is performed using the mean, median, or machine learning-based predictive imputation methods, based on the data distribution characteristics. Outliers are detected and corrected using box plot analysis and the 3σ principle. Subsequently, the data is standardized to unify the data format and units, such as unifying the monetary unit to RMB yuan and the time format to the ISO 8601 standard. Finally, one-hot encoding, label encoding, and other techniques are used to encode categorical variables, transforming them into numerical data that can be processed by machine learning.

[0042] For unstructured data, pre-trained language models (such as BERT) are used to perform semantic understanding on unstructured data such as news information and contract texts. Named entity recognition technology is used to extract key information such as the transaction parties (company name, legal representative), transaction amount, and transaction time. Relationship extraction technology is used to identify the relationships between the transacting parties (such as holding, shareholding, cooperation) and transaction types (purchase and sale, loan, guarantee), transforming the extracted information into structured feature vectors to support subsequent analysis.

[0043] By using graph databases (such as Neo4j) to construct a knowledge graph of related transactions, data fusion can be achieved. Structured and unstructured data can be integrated into the graph, with enterprises as nodes and transaction relationships, equity relationships, etc. as edges. Nodes are given attributes such as enterprise name, industry type, and credit score, while edges are given attributes such as transaction amount, transaction time, and pricing method.

[0044] It should be noted that knowledge graphs enable the deep integration of structured and unstructured data, providing an intuitive display of the complex network of relationships between enterprises.

[0045] Augmentation of the fused data using generative adversarial networks and adversarial training techniques includes:

[0046] Structured data augmentation: Generative adversarial network (GAN) is used to generate simulated transaction data based on the distribution characteristics of the original structured transaction data. By adjusting the parameters of the generator and discriminator of the GAN, it is ensured that the generated data has similar statistical characteristics to the real data, thereby expanding the number of training samples and improving the model's adaptability to different transaction scenarios.

[0047] Unstructured data augmentation: Performing operations such as synonym replacement, sentence restructuring, and random insertion or deletion of words on text data increases the diversity of the data without changing the semantics of the text. At the same time, adding noise (such as random characters or disordered words) simulates interference factors in the data collection process and improves the robustness of the model.

[0048] S2. Spatiotemporal Graph Neural Network Construction and Feature Extraction: A dynamic graph network is constructed with enterprises as nodes and related transaction relationships as edges. Each edge is assigned attributes such as transaction amount, transaction time, and pricing method. The graph is divided according to time windows. A spatial convolutional layer is constructed using a graph attention mechanism to learn local structural features. A temporal convolutional layer is constructed using a temporal convolutional network to extract time series features. The features extracted by the spatial and temporal convolutional layers are fused through a feature fusion layer. The comprehensive risk feature representation of the nodes is output through a fully connected layer.

[0049] Specifically, time windows are divided according to days, weeks, and months; for example, daily transaction data is constructed as a sub-graph, and through the connection of time dimensions, a spatiotemporal graph containing time dimensions is formed to reflect the evolution characteristics of the transaction network over time.

[0050] The graph attention mechanism (GAM) is adopted to adaptively allocate attention weights based on the strength of association between nodes and transaction characteristics. During the calculation process, the multi-head attention mechanism captures the relationship between nodes from multiple perspectives, learns the local structural features of nodes in the transaction network, and explores the implicit associations between nodes. For example, nodes that have frequent transactions with high-risk enterprises are given higher attention weights to highlight their importance in risk propagation.

[0051] Temporal Convolutional Network (TCN) is used to process graph structures across different time windows. Through causal convolution and dilated convolution operations, TCN effectively extracts the time-series features of transaction data, capturing the dynamic trends of transaction risk over time. For example, it identifies periodic fluctuations in transaction frequency and the evolution of risk events over time, providing temporal information support for risk prediction.

[0052] The features extracted by the spatial and temporal convolutional layers are fused, and the comprehensive risk feature representation of the output node is obtained through the fully connected layer. During the fusion process, a weighted summation or gating mechanism is used to assign weights according to the importance of the features, so as to provide comprehensive and rich feature information for subsequent risk assessment.

[0053] S3. Interpretable training of hybrid deep learning models: Construct a hybrid deep learning model that integrates spatiotemporal graph neural networks and Transformers; calculate the loss of the hybrid deep learning model using a weighted cross-entropy loss function combined with a contrastive learning loss function; use the AdamW optimizer to adjust the learning rate and weight decay parameters of the hybrid deep learning model; combine SHAP values ​​and LIME techniques to interpret the model's decision-making process.

[0054] It should be noted that the Spatiotemporal Graph Neural Network (ST-GNN) is responsible for learning the topology and spatiotemporal dynamics of the transaction network, capturing the local relationships and temporal changes between nodes; the Transformer uses a multi-head attention mechanism to capture the long-distance dependencies between transaction features, such as the correlation between the transaction behaviors of different companies at different time periods. The combination of the two enables a comprehensive and in-depth understanding of the inherent patterns of related transaction data.

[0055] Specifically, the weighted cross-entropy loss function assigns different weights to different categories based on the severity of the transaction risk. For example, it assigns higher weights to high-risk transaction categories to improve the model's ability to identify high-risk transactions. The contrastive learning loss function enhances the model's ability to distinguish transaction features by maximizing the feature similarity of similar samples and minimizing the feature similarity of dissimilar samples, enabling the model to accurately identify transaction patterns with different risk levels.

[0056] The AdamW optimizer is used to train the model. By adjusting the learning rate and weight decay parameters, the training speed and generalization performance of the model are balanced. During the training process, a learning rate decay strategy is adopted to gradually reduce the learning rate as the number of training rounds increases, so as to avoid the model from oscillating in the later stage of training, prevent overfitting, and ensure that the model has good performance on both the training set and the test set.

[0057] By combining SHAP values ​​and LIME technology, the decision-making process of the model is explained. SHAP values ​​quantify the contribution of each feature to the risk prediction result from a global perspective. For example, it is calculated that the deviation of the transaction amount contributes 60% to the risk score of a certain transaction. LIME generates an interpretable linear model within a local scope, intuitively demonstrating the basis for the model's risk judgment for specific transaction cases, such as "the counterparty's credit score is <600 and the transaction amount accounts for more than 20% of the company's net assets, so it is judged as high risk." This makes the model's decision results highly interpretable, making it easy for corporate managers and regulatory agencies to understand and use.

[0058] It should be noted that the SHAP value of feature j is calculated using the Kernel SHAP algorithm, with the formula: φ j =Ε[f(X)|X j =x j ]-Ε[f(X)], where φ j Let f(X) represent the SHAP value of feature j, i.e., the global contribution, which measures the impact of feature j on the prediction result. Let f(X) represent the prediction function of the model, and E[·] represent the expected value. j =x j The value of a fixed feature j is represented as x. j For example, "Transaction amount deviation = 30%".

[0059] S4. Real-time risk assessment and dynamic early warning: Input the real-time collected related transaction data into the trained hybrid deep learning model, and use the streaming computing framework Flink to realize real-time data processing and fast model inference, output the risk probability value of the transaction in a short time, and realize the real-time assessment of the risk of related transactions.

[0060] Based on historical trading data and the current market environment, a reinforcement learning algorithm is used to dynamically adjust the risk assessment threshold. An intelligent agent is set up to automatically adjust the threshold based on the risk assessment results and feedback from actual risk events, such as whether a real risk has occurred and whether the warning is accurate. Through continuous trial and error learning, the agent can adjust the threshold automatically. For example, when the market environment is highly volatile, the agent lowers the threshold to improve the sensitivity of risk warnings. When the false alarm rate is too high, the agent raises the threshold to reduce unnecessary warnings, thus balancing the false alarm rate and the missed alarm rate, making risk warnings more accurate and timely.

[0061] When the probability of a transaction's risk exceeds a dynamically adjusted threshold, the system automatically generates detailed risk warning information. This warning includes basic information about the risky transaction, its risk level (high, medium, low), an explanation of its risk characteristics, and the potential scope of its impact. The basic information includes the transacting parties, transaction amount, transaction time, and transaction type. The risk level is categorized as high, medium, and low. The risk characteristic explanation is based on SHAP and LIME analysis results. The potential scope of its impact is analyzed using a knowledge graph to identify related enterprises. The warning information is pushed to relevant managers through multiple channels, including email, SMS, and internal management systems. A visual interface prominently displays the risky transaction's location and propagation path within the related transaction network, helping managers quickly understand the risk situation and take timely countermeasures.

[0062] S5. Risk Transmission Prediction and Intelligent Decision-Making: Based on a trained model and a spatiotemporal diagram of related transactions, Monte Carlo simulation is used to predict the transmission path and scope of risk in the transaction network. Specifically, through multiple random simulations of the risk diffusion process, considering factors such as transaction amount, transaction frequency, and corporate credit status, the probability and potential loss of different nodes being affected by risk at different time points are calculated. For example, simulating how risk spreads to other enterprises through the related transaction network after a core enterprise defaults, identifying key nodes and vulnerable links in risk transmission, and providing a scientific basis for risk prevention and control decisions;

[0063] A deep reinforcement learning-based intelligent decision-making model is constructed, using risk transmission prediction results as state input and control measures as actions, such as suspending transactions, restricting capital flow, requiring supplementary collateral, and initiating internal audits. The objective functions are minimizing risk loss and minimizing control costs. The intelligent decision-making model continuously optimizes its strategies through interactive learning with the environment. For example, when facing high-risk transactions, the model can quickly evaluate the effectiveness of different control measures and select the optimal solution that effectively controls risk while reducing corporate losses, thus achieving automated and intelligent risk control decision-making.

[0064] S6. Continuous Model Evolution and Adaptation: A concept drift detection method based on Dynamic Time Warping (DTW) and cluster analysis is adopted to monitor the distribution changes and pattern evolution of transaction data in real time. DTW is used to measure the similarity between time series data. By comparing the pattern differences between current data and historical data, it is determined whether concept drift has occurred. Cluster analysis divides the data into different categories and observes the changes in the distribution of data in each category. When a significant change in data distribution is detected (such as the emergence of new trading patterns or changes in risk characteristics), the model update mechanism is triggered to promptly detect signs of model performance degradation.

[0065] Specifically, obtain the historical transaction dataset A. hist Real-time transaction flow Areal-time And divided into D according to the time window. t D t+1 ..., the historical transaction dataset is split according to transaction type, and the time-series feature sequences of each type are extracted. For real-time window D t Extracting time series sequences of the same type Calculated using DTW and The normalization distance is calculated and compared with a preset normalization distance threshold. If the normalization distance is greater than the preset normalization distance threshold, the transaction type is determined to be time-series drift.

[0066] For A hist Extract transaction features and use the DBSCAN algorithm to cluster them into K clusters. And record the density of each cluster. center For D t Using the same features and clustering parameters, clusters are obtained. Calculate density change Center offset It is then compared with preset density change thresholds and center offset thresholds. If the number of clusters exceeding the thresholds is greater than K / 2, then feature distribution drift is determined.

[0067] If either time-series drift or feature drift is triggered, and the proportion of drift-covered transactions exceeds a preset threshold, then the model update mechanism is triggered to mark the drift type.

[0068] It should be noted that the process of calculating the normalized distance using DTW is as follows:

[0069] Obtain historical time series and real-time time series Calculate the Euclidean distance between each point in the two sequences, generating an n×m distance matrix D, where

[0070] Starting from the top left corner D[1][1] of matrix D, move to the bottom right corner D[n][m]. Each step can only be right, down, or diagonally. Find a path P such that the sum of distances along the path is minimized (min∑). (i,j∈P) D[i][j], and the total distance of this path is the DTW normalized distance d. dtw .

[0071] Incremental learning techniques are used to gradually integrate new data into the model training process. In incremental learning, the model learns from new data while retaining existing knowledge, avoiding the "catastrophic forgetting" phenomenon. Simultaneously, transfer learning techniques are combined to transfer learned knowledge and features to similar trading scenarios or industry data. For example, a risk identification model trained in one industry can be transferred to other industries with similar trading patterns, accelerating the model's adaptability to new data and reducing the time and amount of data required for model training. Regular comprehensive evaluation and optimization of the model are conducted, and risk assessment strategies are updated based on the evaluation results to ensure that the model maintains good performance and accuracy in constantly changing trading environments.

[0072] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. An analytical method for analyzing the risk of related-party transactions within a group using machine learning, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data fusion and enhancement: Comprehensive collection of related transaction data from within and outside the enterprise group; preprocessing of structured data using feature engineering techniques; semantic understanding and information extraction of unstructured data using pre-trained language models; data fusion by constructing a related transaction knowledge graph using graph databases; and enhancement processing of the fused data using generative adversarial networks and adversarial training techniques. S2. Spatiotemporal Graph Neural Network Construction and Feature Extraction: A dynamic graph network is constructed with enterprises as nodes and related transaction relationships as edges. A spatiotemporal graph is formed according to the time window. A spatiotemporal graph neural network is designed. The local structural features of the nodes are learned through the graph attention mechanism of the spatial convolutional layer. The temporal convolutional layer uses the temporal convolutional network to extract time series features. The feature fusion layer fuses the two features and outputs a comprehensive risk feature representation of the nodes. S3. Interpretable Hybrid Deep Learning Model Training: Construct a hybrid deep learning model that integrates spatiotemporal graph neural networks and Transformers. Optimize the loss using a weighted cross-entropy loss function combined with a contrastive learning loss function. Train the model using the AdamW optimizer and enhance model interpretability by combining SHAP values ​​and LIME techniques. S4. Real-time risk assessment and dynamic early warning: Input the real-time collected related transaction data into the trained model, use the streaming computing framework to achieve real-time processing and inference, and dynamically adjust the risk assessment threshold based on historical data and market environment using reinforcement learning algorithms. When the risk probability exceeds the threshold, risk early warning information is automatically generated and pushed through multiple channels. S5. Risk Transmission Prediction and Intelligent Decision-Making: Based on the trained model and spatiotemporal map, the Monte Carlo simulation method is used to predict the risk transmission path and impact range. An intelligent decision-making model based on deep reinforcement learning is constructed to automatically generate the optimal risk prevention and control strategy with the goal of minimizing risk loss and prevention and control costs. S6. Continuous Model Evolution and Adaptation: A concept drift detection method based on dynamic time warping and cluster analysis is used to monitor data changes. Incremental learning and transfer learning techniques are used to update the model, and the model is evaluated and optimized regularly.

2. The analytical method for analyzing the risk of related-party transactions of a group through machine learning according to claim 1, characterized in that, The preprocessing of structured data using feature engineering techniques includes data cleaning, standardization, and feature encoding. The unstructured data processing extracts key transaction information from the text using natural language processing (NLP) technology and transforms it into structured feature vectors. NLP technology includes, but is not limited to, named entity recognition (RANK) technology and relation extraction technology. Key transaction information includes, but is not limited to, transaction subject, transaction amount, transaction time, and risk keywords. The construction of the related-party transaction knowledge graph involves using enterprises as nodes and related-party transactions as edges, constructing the related-party transaction knowledge graph using a graph database, assigning attributes to nodes and edges, where related-party transaction relationships include transaction relationships and equity relationships, assigning attributes to nodes including enterprise name, industry type, and credit score, and assigning attributes to edges including transaction amount, transaction time, and pricing method.

3. The analytical method for analyzing the risk of related-party transactions of a group through machine learning according to claim 1, characterized in that, The time window is divided into days, weeks, and months; The spatial convolutional layer adaptively allocates attention weights based on the correlation strength and transaction characteristics between nodes; The temporal convolutional layer captures the dynamic changes in trading risk over time, including the periodic fluctuations in trading frequency and the evolution patterns of risk events.

4. The analytical method for analyzing the risk of related-party transactions of a group through machine learning according to claim 1, characterized in that, The weighted cross-entropy loss function assigns different weights to different categories based on the severity of the trading risk; The SHAP value quantifies the contribution of each feature to the risk prediction result from a global perspective; The LIME generates interpretable linear models within a local scope.

5. The analytical method for analyzing the risk of related-party transactions within a group using machine learning according to claim 1, characterized in that, The risk warning information includes basic information about the risky transaction, risk level, explanation of risk characteristics, and the possible scope of risk impact. It also displays the location of the risky transaction in the related transaction network and the risk propagation path on a visual interface.

6. The analytical method for analyzing the risk of related-party transactions of a group through machine learning according to claim 1, characterized in that, The Monte Carlo simulation method calculates the probability and potential loss of different nodes being affected by the risk at different time points by repeatedly simulating the risk diffusion process. The intelligent decision-making model takes the risk transmission prediction results as the state input and the prevention and control measures as the actions, and generates the optimal risk prevention and control strategy with the goal of minimizing risk loss and minimizing prevention and control costs.

7. The analytical method for analyzing the risk of related-party transactions of a group through machine learning according to claim 1, characterized in that, The concept drift detection method monitors the distribution changes and pattern evolution of transaction data in real time, and triggers a model update mechanism when a significant change in data distribution is detected; the incremental learning technology gradually integrates new data into the model training process, and the transfer learning technology transfers learned knowledge and features in similar transaction scenarios or industry data.

8. The analytical method for analyzing the risk of related-party transactions of a group through machine learning according to claim 7, characterized in that, The execution process of the triggering model update mechanism is as follows: Obtain historical transaction dataset A hist Real-time transaction flow A real-time And divided into D according to the time window. t D t+1 ..., the historical transaction dataset is split according to transaction type, and the time-series feature sequences of each type are extracted. For real-time window D t Extracting time series sequences of the same type Calculated using DTW and The normalization distance is calculated and compared with a preset normalization distance threshold. If the normalization distance is greater than the preset normalization distance threshold, the transaction type is determined to be time-series drift. For A hist Extract transaction features and use the DBSCAN algorithm to cluster them into K clusters. And record the density of each cluster. center For D t Using the same features and clustering parameters, clusters are obtained. And record the density of each cluster. center Calculate density change Center offset It is then compared with preset density change thresholds and center offset thresholds. If the number of clusters exceeding the thresholds is greater than K / 2, then feature distribution drift is determined. If either time-series drift or feature drift is triggered, and the proportion of drift-covered transactions exceeds a preset threshold, then the model update mechanism is triggered to mark the drift type.

Images