Digital economy transaction processing and anomaly detection methods based on machine learning

Abstract

The present invention relates to the field of digital information processing technology, and in particular to a method for processing and detecting digital economic transactions based on machine learning. The method comprises the following steps: collecting digital transaction data and performing weighted feature analysis on a digital economic transaction platform to generate weighted digital transaction feature data; analyzing digital transaction strategy prediction data based on the weighted digital transaction feature data; performing abnormal transaction confrontation game feature analysis based on the digital transaction strategy prediction data to generate abnormal transaction confrontation game feature data; establishing a digital transaction fraud tracing tree model based on the abnormal transaction confrontation game feature data; performing digital transaction abnormal feedback data analysis on the weighted digital transaction feature data based on the digital transaction fraud tracing tree model to generate digital transaction abnormal feedback data. The present invention realizes efficient detection of digital economic transaction anomalies by predicting the changeable and complex fraudulent transaction behaviors in digital transaction behaviors.

Description

Technical Field

[0001] The present invention relates to the field of digital information processing technology, and in particular to a digital economic transaction processing and anomaly detection method based on machine learning. Background Art

[0002] With the rapid development of the digital economy, transaction activity in areas such as e-commerce, online payments, and digital finance has increased significantly. This has been accompanied by a surge in transaction data and its complexity, often containing diverse information such as transaction amounts, time, location, participant information, product descriptions, and user behavior. Due to the diversity and complexity of transaction behavior, fraudulent methods in the digital economy are constantly evolving. To evade anomaly detection and tracking of fraudulent methods in the digital economy, intelligent technologies are needed to continuously detect anomalies in digital economy transactions and ensure a healthy digital economy trading environment. However, traditional digital economy transaction processing and anomaly detection methods rely on predefined rules or thresholds to execute anomalous digital transaction detection methods. These methods are based on past experience and known fraud patterns. As fraudulent methods continue to evolve, fraudsters are able to devise increasingly complex and covert anomalous behaviors to evade traditional anomalous digital transaction detection methods, making it difficult to adapt to the changing nature of digital transactions. Summary of the Invention

[0003] Based on this, the present invention provides a digital economic transaction processing and anomaly detection method based on machine learning to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a method for digital economic transaction processing and anomaly detection based on machine learning is provided, comprising the following steps:

[0005] Step S1: collecting digital transaction multimodal data from a digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain digital transaction multimodal data; performing digital transaction feature analysis on the digital transaction multimodal data to generate digital transaction feature data;

[0006] Step S2: Designing digital transaction attention weight information based on the digital transaction multimodal data; performing attention weighting processing on the digital transaction feature data according to the digital transaction attention weight information to generate weighted digital transaction feature data;

[0007] Step S3: Acquire historical digital transaction strategies; establish a digital transaction strategy reinforcement learning model based on the historical digital transaction strategies; transmit the weighted digital transaction feature data to the digital transaction strategy reinforcement learning model for digital transaction strategy prediction processing to generate digital transaction strategy prediction data;

[0008] Step S4: performing a multi-dimensional digital transaction fraud feature derivative analysis based on the digital transaction strategy prediction data to generate multi-dimensional digital transaction fraud feature derivative data; designing an abnormal transaction adversarial game mathematical model based on the multi-dimensional digital transaction fraud feature derivative data to generate an abnormal transaction adversarial game model;

[0009] Step S5: Analyzing and processing abnormal transaction adversarial game characteristics based on the abnormal transaction adversarial game model to generate abnormal transaction adversarial game characteristic data; establishing a preliminary digital transaction fraud tracing tree model based on the abnormal transaction adversarial game characteristic data; optimizing the tree structure weight parameters of the preliminary digital transaction fraud tracing tree model based on the digital transaction attention weight information corresponding to the weighted digital transaction characteristic data to obtain the digital transaction fraud tracing tree model;

[0010] Step S6: Analyze and process the weighted digital transaction feature data based on the digital transaction fraud tracing tree model to generate digital transaction abnormality feedback data; transmit the digital transaction abnormality feedback data to the terminal to execute the digital transaction abnormality feedback operation.

[0011] The present invention uses a pre-configured digital economic transaction monitoring engine to collect multimodal digital transaction data, ensuring that the collected data covers all key aspects of the trading platform, including user behavior, transaction time, transaction amount, and other dimensions. This multimodal data collection improves data comprehensiveness and helps capture complex transaction patterns and potential abnormal behaviors. Digital transaction feature analysis and processing of the collected multimodal data ensures the effective extraction of key features from the data, laying a solid foundation for subsequent analysis and processing, and helping to improve the accuracy of anomaly detection. Digital transaction attention weight information is designed based on multimodal digital transaction data, and attention weighting is applied to digital transaction feature data based on this information. This can highlight specific important features in the transaction data, thereby improving the learning effect of subsequent models on these key features and enhancing the accuracy and reliability of predictions. Historical digital transaction strategies are obtained and, based on this data, a digital transaction strategy reinforcement learning model is established, effectively leveraging the lessons learned from historical data to optimize the current trading strategy model. The weighted digital transaction feature data is transmitted to the reinforcement learning model for strategy prediction. The model can predict the behavioral combinations of digital transactions based on the weighted digital transaction feature data, improving the effectiveness of trading decisions. Based on digital transaction strategy prediction data, multi-dimensional digital transaction fraud feature derivation analysis and processing are performed to deeply explore potential fraud behavior patterns, providing a rich fraud feature foundation for subsequent anomaly detection and ensuring that the anomaly detection model can identify complex fraud behaviors. Designing an abnormal transaction adversarial game mathematical model based on this multi-dimensional digital transaction fraud feature derivation data can improve the robustness of digital transaction anomaly detection, making it more resistant to abnormal transaction fraud and thus more effective in combating complex fraud behaviors. The abnormal transaction adversarial game feature analysis and processing performed using the abnormal transaction adversarial game model dynamically analyzes the behavior patterns of different participants in the transaction process. The generated abnormal transaction adversarial game feature data provides an important basis for tracing fraud behaviors. Based on the abnormal transaction adversarial game feature data and attention weight information, the preliminary digital transaction fraud tracing tree model is gradually optimized, resulting in a more accurate and efficient fraud tracing tree model. This tree model can efficiently trace and identify abnormal transactions in complex transaction networks, providing strong support for fraud detection in the digital economy. Based on the optimized digital transaction fraud tracing tree model, the weighted digital transaction feature data is analyzed and processed to generate effective digital transaction anomaly feedback data. After being transmitted to the terminal, these feedback data can be used to implement feedback measures for abnormal transactions, thereby forming a closed-loop feedback system, continuously improving and optimizing the transaction anomaly detection mechanism, and improving the security and reliability of the entire system.

[0012] The present invention provides a beneficial effect in that the machine learning-based digital economic transaction processing and anomaly detection method of the present invention, by introducing a machine learning model, particularly a reinforcement learning model, can adaptively adjust digital transaction strategies and detection methods. Unlike traditional methods that rely on predefined rules and thresholds, the machine learning model extracts features from historical data and predicts future transaction behavior through continuous learning and optimization. This dynamic adaptability enables the system to promptly detect and respond to emerging fraud patterns, no longer limited to past experience and known fraud patterns. By designing digital transaction attention weight information and adversarial game mathematical models, the system can handle more complex and hidden anomalies. The weighted processing of attention weights enables the system to focus resources on analyzing anomalous transaction features, improving the accuracy and efficiency of anomaly detection. The introduction of adversarial game models enables the system to simulate and counter potential fraud strategies and identify complex fraud behaviors that circumvent traditional detection methods. Furthermore, during the anomaly detection process, the model is continuously optimized. By optimizing the weight parameters of the preliminary digital transaction fraud tracing tree model, the model's accuracy and effectiveness are continuously improved. The continuous learning and optimization capabilities keep pace with evolving fraud behaviors, gradually enhancing the detection capabilities of new and complex digital transaction anomalies. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flowchart of the steps of a method for digital economic transaction processing and anomaly detection based on machine learning in the present invention;

[0014] Figure 2 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0015] Figure 3 for Figure 1 Detailed implementation steps of step S5;

[0016] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0017] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0018] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0019] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0020] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for digital economic transaction processing and anomaly detection based on machine learning, comprising the following steps:

[0021] Step S1: collecting digital transaction multimodal data from a digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain digital transaction multimodal data; performing digital transaction feature analysis on the digital transaction multimodal data to generate digital transaction feature data;

[0022] Step S2: Designing digital transaction attention weight information based on the digital transaction multimodal data; performing attention weighting processing on the digital transaction feature data according to the digital transaction attention weight information to generate weighted digital transaction feature data;

[0023] Step S3: Acquire historical digital transaction strategies; establish a digital transaction strategy reinforcement learning model based on the historical digital transaction strategies; transmit the weighted digital transaction feature data to the digital transaction strategy reinforcement learning model for digital transaction strategy prediction processing to generate digital transaction strategy prediction data;

[0024] Step S4: performing a multi-dimensional digital transaction fraud feature derivative analysis based on the digital transaction strategy prediction data to generate multi-dimensional digital transaction fraud feature derivative data; designing an abnormal transaction adversarial game mathematical model based on the multi-dimensional digital transaction fraud feature derivative data to generate an abnormal transaction adversarial game model;

[0025] Step S5: Analyzing and processing abnormal transaction adversarial game characteristics based on the abnormal transaction adversarial game model to generate abnormal transaction adversarial game characteristic data; establishing a preliminary digital transaction fraud tracing tree model based on the abnormal transaction adversarial game characteristic data; optimizing the tree structure weight parameters of the preliminary digital transaction fraud tracing tree model based on the digital transaction attention weight information corresponding to the weighted digital transaction characteristic data to obtain the digital transaction fraud tracing tree model;

[0026] Step S6: Analyze and process the weighted digital transaction feature data based on the digital transaction fraud tracing tree model to generate digital transaction abnormality feedback data; transmit the digital transaction abnormality feedback data to the terminal to execute the digital transaction abnormality feedback operation.

[0027] The present invention uses a pre-configured digital economic transaction monitoring engine to collect multimodal digital transaction data, ensuring that the collected data covers all key aspects of the trading platform, including user behavior, transaction time, transaction amount, and other dimensions. This multimodal data collection improves data comprehensiveness and helps capture complex transaction patterns and potential abnormal behaviors. Digital transaction feature analysis and processing of the collected multimodal data ensures the effective extraction of key features from the data, laying a solid foundation for subsequent analysis and processing, and helping to improve the accuracy of anomaly detection. Digital transaction attention weight information is designed based on multimodal digital transaction data, and attention weighting is applied to digital transaction feature data based on this information. This can highlight specific important features in the transaction data, thereby improving the learning effect of subsequent models on these key features and enhancing the accuracy and reliability of predictions. Historical digital transaction strategies are obtained and, based on this data, a digital transaction strategy reinforcement learning model is established, effectively leveraging the lessons learned from historical data to optimize the current trading strategy model. The weighted digital transaction feature data is transmitted to the reinforcement learning model for strategy prediction. The model can predict the behavioral combinations of digital transactions based on the weighted digital transaction feature data, improving the effectiveness of trading decisions. Based on digital transaction strategy prediction data, multi-dimensional digital transaction fraud feature derivation analysis and processing are performed to deeply explore potential fraud behavior patterns, providing a rich fraud feature foundation for subsequent anomaly detection and ensuring that the anomaly detection model can identify complex fraud behaviors. Designing an abnormal transaction adversarial game mathematical model based on this multi-dimensional digital transaction fraud feature derivation data can improve the robustness of digital transaction anomaly detection, making it more resistant to abnormal transaction fraud and thus more effective in combating complex fraud behaviors. The abnormal transaction adversarial game feature analysis and processing performed using the abnormal transaction adversarial game model dynamically analyzes the behavior patterns of different participants in the transaction process. The generated abnormal transaction adversarial game feature data provides an important basis for tracing fraud behaviors. Based on the abnormal transaction adversarial game feature data and attention weight information, the preliminary digital transaction fraud tracing tree model is gradually optimized, resulting in a more accurate and efficient fraud tracing tree model. This tree model can efficiently trace and identify abnormal transactions in complex transaction networks, providing strong support for fraud detection in the digital economy. Based on the optimized digital transaction fraud tracing tree model, the weighted digital transaction feature data is analyzed and processed to generate effective digital transaction anomaly feedback data. After being transmitted to the terminal, these feedback data can be used to implement feedback measures for abnormal transactions, thereby forming a closed-loop feedback system, continuously improving and optimizing the transaction anomaly detection mechanism, and improving the security and reliability of the entire system.

[0028] As an embodiment of the present invention, refer to Figure 1The above is a schematic flow chart of the steps of a method for processing and detecting digital economic transactions based on machine learning according to the present invention. In this embodiment, the method for processing and detecting digital economic transactions based on machine learning includes the following steps:

[0029] Step S1: collecting digital transaction multimodal data from a digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain digital transaction multimodal data; performing digital transaction feature analysis on the digital transaction multimodal data to generate digital transaction feature data;

[0030] In an embodiment of the present invention, a digital economic transaction monitoring engine is deployed on a digital economic transaction platform, such as an online e-commerce platform or an electronic transaction platform. The engine is capable of real-time monitoring and collecting transaction information from multiple data sources, including user behavior data (such as clickstreams, browsing history), transaction data (such as order amounts, payment methods), time and geographic location data, etc. Through the monitoring engine, multimodal data collection is performed on all transaction activities on the platform. The collected data includes structured data (such as transaction amounts, user IDs) and unstructured data (such as product descriptions and user reviews). The collected data is uniformly stored in a data warehouse for subsequent analysis. The collected multimodal data is preprocessed, including data cleaning, deduplication, and formatting. Feature extraction is performed on the cleaned data. Corresponding data feature methods are designed according to different transaction behavior patterns to extract features such as user purchase frequency, transaction time interval, and transaction amount volatility. These features are preliminarily analyzed using a data clustering algorithm to identify typical transaction patterns and behavior features, and are converted into structured feature data to generate digital transaction feature data for subsequent steps.

[0031] Step S2: Designing digital transaction attention weight information based on the digital transaction multimodal data; performing attention weighting processing on the digital transaction feature data according to the digital transaction attention weight information to generate weighted digital transaction feature data;

[0032] In an embodiment of the present invention, the self-attention mechanism in deep learning is used to analyze the generated digital transaction feature data and design digital transaction attention weight information. The attention mechanism can identify the most influential features in the transaction data, such as the time period of high-frequency transactions, the abnormal consumption behavior of specific users, and the consumption frequency, etc., and design corresponding attention weight information. The attention weight information is designed based on the relevant information that is most likely to cause digital transaction anomalies, and different weight values ​​are assigned to these features. According to the attention weight information, each feature data is weighted, and the weight coefficient corresponding to each feature data is identified to highlight the most important features. The weighted feature data will be able to provide more accurate results in subsequent model training and prediction, and generate weighted digital transaction feature data.

[0033] Step S3: Acquire historical digital transaction strategies; establish a digital transaction strategy reinforcement learning model based on the historical digital transaction strategies; transmit the weighted digital transaction feature data to the digital transaction strategy reinforcement learning model for digital transaction strategy prediction processing to generate digital transaction strategy prediction data;

[0034] In an embodiment of the present invention, previous transaction strategies are extracted from historical transaction data and archived. Historical strategies include payment methods, purchasing behavior patterns, preferential usage, etc. adopted by users in different situations. A digital transaction strategy model is constructed using a reinforcement learning algorithm (such as deep Q learning). The state space of the model includes the user's behavioral characteristics, transaction environment characteristics, etc., and the action space includes transaction decisions that can be taken, such as recommending different products, adjusting payment methods, etc. The historical transaction strategies are input into the model for training. The model gradually learns and optimizes its strategy selection by simulating the execution effects of different strategies in different situations. The weighted digital transaction feature data is input into the trained reinforcement learning model, and the model makes strategy predictions based on the current transaction status. The output of the model is a transaction prediction strategy for the current transaction scenario, such as recommending a certain payment method, triggering a certain preferential measure, etc. The generated strategy prediction data will be used as a basis and input into the next step for derivative analysis of fraud features.

[0035] Step S4: performing a multi-dimensional digital transaction fraud feature derivative analysis based on the digital transaction strategy prediction data to generate multi-dimensional digital transaction fraud feature derivative data; designing an abnormal transaction adversarial game mathematical model based on the multi-dimensional digital transaction fraud feature derivative data to generate an abnormal transaction adversarial game model;

[0036] In this embodiment of the present invention, based on digital transaction strategy prediction data, Generative Adversarial Network (GAN) technology is used to perform multi-dimensional fraud feature derivation analysis. By analyzing user transaction behavior and predicted strategy data, existing fraud pattern characteristics are extracted. For example, high-frequency small-value transactions and abnormal consumption within a specific time period.

[0037] Using Generative Adversarial Network (GAN) technology, these features are combined and transformed to generate derivative data that can cover a variety of potential abnormal transaction patterns. An abnormal transaction adversarial game model is constructed based on the derived data of multi-dimensional digital transaction fraud features. The abnormal transaction adversarial game model simulates fraudulent behaviors encountered in transactions and the corresponding response strategies. By continuously learning the derived fraudulent behaviors and corresponding fraud response strategies from the derived data of multi-dimensional digital transaction fraud features, the game model optimizes the anomaly detection strategy, enabling the system to identify more complex and hidden fraudulent behaviors. The generated abnormal transaction adversarial game model includes digital transaction fraudulent behaviors and the corresponding fraud response strategies, which will be further used for subsequent feature analysis and traceability tree model design.

[0038] Step S5: Analyzing and processing abnormal transaction adversarial game characteristics based on the abnormal transaction adversarial game model to generate abnormal transaction adversarial game characteristic data; establishing a preliminary digital transaction fraud tracing tree model based on the abnormal transaction adversarial game characteristic data; optimizing the tree structure weight parameters of the preliminary digital transaction fraud tracing tree model based on the digital transaction attention weight information corresponding to the weighted digital transaction characteristic data to obtain the digital transaction fraud tracing tree model;

[0039] In an embodiment of the present invention, the abnormal transaction behavior feature data generated in the game model is analyzed to extract key adversarial features. These features represent the strategies adopted by fraudsters to evade detection, such as dispersing transaction amounts and operating in different locations.

[0040] Based on the extracted adversarial features, a preliminary digital transaction fraud tracing tree model was designed. The model's nodes represent different fraud behavior characteristics, and the paths represent the correlations between the characteristics. The digital transaction attention weight information corresponding to the weighted digital transaction feature data was combined with the adversarial feature data to optimize the weight parameters of the tracing tree model and adjust the weights of each node so that the model can more accurately reflect the evolutionary path of actual fraudulent behavior. By gradually adjusting and optimizing the preliminary tracing tree model, a tuned digital transaction fraud tracing tree model was ultimately obtained. This model can dynamically adjust the priority of nodes and paths based on the weight information of different features, ensuring that fraudulent behavior can be quickly identified and traced when it occurs.

[0041] Step S6: Analyze and process the weighted digital transaction feature data based on the digital transaction fraud tracing tree model to generate digital transaction abnormality feedback data; transmit the digital transaction abnormality feedback data to the terminal to execute the digital transaction abnormality feedback operation.

[0042] In an embodiment of the present invention, an optimized digital transaction fraud tracing tree model is used to perform anomaly prediction processing on weighted digital transaction feature data. Based on the distribution and correlation of the weighted digital transaction feature data, the model predicts the abnormal risk level of the current transaction. For example, the model analyzes the risks encountered in current and future transaction decisions based on the current transaction feature data to generate digital transaction anomaly prediction data. The digital transaction anomaly prediction data is compared with pre-set anomaly assessment criteria to determine the accuracy of the anomaly prediction. During the assessment process, the system considers various factors, such as abnormal fluctuations in transaction amounts and abnormal changes in user behavior. These factors are converted into quantitative anomaly indicators based on the assessment criteria to generate final anomaly assessment data. The anomaly assessment data is mapped and integrated with the predicted data to generate comprehensive anomaly feedback data, summarizing the analysis results of digital transaction anomalies and corresponding quantitative indicators, and providing detailed explanations and recommendations for predicted abnormal behavior. The generated anomaly feedback data is transmitted to the terminal system for further action or automated response. Based on the feedback information, the terminal can implement pre-set security measures, such as freezing accounts, restricting transactions, and sending alerts, to ensure the security and stability of the transaction environment.

[0043] Preferably, step S1 includes the following steps:

[0044] Step S11: collecting digital transaction multimodal data from the digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain digital transaction multimodal data;

[0045] Step S12: performing digital transaction behavior analysis based on the digital transaction multimodal data to generate digital transaction behavior data;

[0046] Step S13: performing cluster analysis on the digital transaction behavior data to generate clustered digital transaction behavior data;

[0047] Step S14: performing transaction pattern recognition processing based on the clustered digital transaction behavior data to generate digital transaction pattern data;

[0048] Step S15: Designing a transaction pattern dimensionality reduction feature matrix based on the digital transaction pattern data to generate a transaction pattern dimensionality reduction feature matrix;

[0049] Step S16: performing digital transaction feature mapping processing on the digital transaction multimodal data according to the transaction mode dimension reduction feature matrix to generate digital transaction feature data.

[0050] The present invention uses a pre-set digital economic transaction monitoring engine to collect multimodal data from digital economic transaction platforms, enabling comprehensive and accurate acquisition of transaction data across different dimensions. This provides a rich and diverse information source for subsequent data analysis, helping to improve the comprehensiveness and reliability of anomaly detection. Behavioral analysis of multimodal digital transaction data helps uncover the underlying characteristics of user transaction behavior. This analysis can identify implicit behavioral patterns, improve the ability to identify abnormal transaction behavior, and more accurately capture abnormal signals. Cluster analysis of digital transaction behavior data effectively groups transaction behavior data and identifies different transaction patterns. The clustering results provide more accurate input for subsequent pattern recognition and feature extraction, thereby improving the accuracy of anomaly detection. Transaction pattern recognition based on the clustering results identifies different behavioral patterns in the transaction data, helping to establish an understanding of different transaction behavior patterns and thus designing dimensionality reduction features for the data under each pattern. By designing a dimensionality reduction feature matrix for transaction pattern data, the data structure is simplified, computational complexity is reduced, and key information is retained. This not only improves the system's processing speed but also enhances the model's sensitivity to abnormal patterns. The reduced-dimensional feature matrix of transaction patterns is mapped to the multimodal data of digital transactions to accurately extract features related to anomaly detection in digital transactions. This mapping process enables the system to better capture abnormal features in transaction behavior, thereby improving the effectiveness and accuracy of detection.

[0051] In an embodiment of the present invention, a preset digital economic transaction monitoring engine is deployed on a digital economic transaction platform. The engine is composed of multiple modules and can monitor and collect various types of transaction data from the platform in real time. The engine is configured to collect multimodal data, including but not limited to the following data types: transaction data (such as order amount, payment method), user behavior data (such as browsing history, click events), time data (such as transaction time, response time), geographic data (such as transaction location, IP address) and social data (such as user comments, interaction frequency). After the monitoring engine starts working, it continuously collects multimodal data for all transaction activities on the platform. The data collection process includes real-time capture of user operations on the platform, detailed information on transaction execution, and auxiliary information related to the transaction. The collected multimodal data is preprocessed, including data cleaning (removing noise and outliers), format standardization (unifying data format and units), data deduplication (eliminating duplicate data) and other steps. The preprocessed data ensures data quality and consistency, providing a reliable foundation for subsequent behavioral analysis to obtain digital transaction multimodal data. Advanced data analysis techniques (such as time series analysis, frequency analysis, and trend analysis) are used to conduct in-depth behavioral analysis of multimodal data, focusing on user transaction behavior patterns, such as purchase frequency, average transaction amount, purchase time preferences, and geographic location changes. By analyzing this behavioral data, typical transaction behaviors of each user or user group are identified and corresponding digital transaction behavior data is generated. Appropriate clustering algorithms, such as K-means, hierarchical clustering, or density-based clustering (DBSCAN), are selected to process the generated digital transaction behavior data. Clustering is performed on the digital transaction behavior data, grouping similar transactions into the same group. For example, users with similar purchase frequency and transaction amounts can be grouped together, or transactions with similar geographic locations and purchase time patterns can be grouped together. The clustering results reflect both differences and commonalities in transaction behavior among platform users, generating clustered digital transaction behavior data that provides input for subsequent transaction pattern recognition. Based on the clustered behavioral data, a transaction pattern recognition model is constructed. This model can identify transaction patterns using machine learning algorithms such as support vector machines (SVMs), decision trees, random forests, or neural networks. By training on various types of clustered behavior data, the model learns the characteristics of different transaction behaviors and can accurately identify specific transaction patterns, such as high-frequency small-value transactions, large single transactions, and transactions with rapidly changing geographic locations. Clustered digital transaction behavior data is input into the pattern recognition model, which analyzes this data and identifies various transaction patterns. For each transaction pattern, the model generates corresponding digital transaction pattern data. This pattern data describes the typical characteristics of each type of transaction behavior. The model can also identify anomalous transaction patterns, providing an important basis for subsequent feature dimensionality reduction and mapping.Based on pattern recognition, various features within the transaction pattern data are selected and reduced in dimensionality. Dimensionality reduction techniques such as principal component analysis (PCA) and linear discriminant analysis (LDA) are used to reduce the high-dimensional transaction pattern feature data into a low-dimensional feature matrix. This dimensionality reduction process retains the most representative features of transaction behavior, such as transaction time, amount, and frequency, while removing redundant or noisy features, ensuring that the model can handle more concise, information-dense data in subsequent processing. A transaction pattern dimensionality reduction feature matrix is ​​designed, containing the core features after dimensionality reduction. These features are arranged and combined in the matrix to facilitate subsequent feature mapping and data processing. The matrix is ​​designed to highlight the features that best reflect the transaction pattern, thereby improving the model's computational efficiency and predictive capabilities. The transaction pattern dimensionality reduction feature matrix is ​​used to perform feature mapping on the original multimodal data, and the key features after dimensionality reduction are mapped back to the multimodal dataset to generate new digital transaction feature data. Feature mapping ensures that the model can focus on analyzing those features identified as the most important in the dimensionality reduction process, ignoring noise and minor features, thereby improving the processing efficiency of subsequent steps. The mapped data generates the final digital transaction feature data, which will have a high information density and representativeness, and is suitable for further analysis and processing, and ultimately generates digital transaction feature data, including feature data with higher value under different transaction patterns.

[0052] Preferably, step S11 includes the following steps:

[0053] Collecting initial digital transaction multimodal data from the digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain initial digital transaction multimodal data;

[0054] performing data cleaning on the initial digital transaction multimodal data to obtain cleaned digital transaction multimodal data;

[0055] Data standardization is performed on the cleaned digital transaction multimodal data to obtain digital transaction multimodal data.

[0056] The present invention collects initial digital transaction multimodal data on the digital economic transaction platform through a preset digital economic transaction monitoring engine, and comprehensively obtains multimodal data on the transaction platform from multiple dimensions (such as user behavior, transaction type, transaction time, etc.), which can capture detailed information in the transaction process to the greatest extent, improve the diversity and coverage of the data, and lay a solid foundation for subsequent analysis. The initial digital transaction multimodal data is cleaned to effectively remove invalid data such as noise, errors and outliers, improve the quality of the data set, ensure that data related to digital transactions can be accurately obtained, thereby improving the reliability of subsequent data processing and analysis, and reducing analysis errors caused by data quality issues. The cleaned digital transaction multimodal data is standardized so that data from different sources and different types can be compared and analyzed on a unified scale, eliminating deviations caused by different data scales, enhancing the consistency and comparability of the data, and providing more reliable input data for subsequent data feature extraction and model training.

[0057] In an embodiment of the present invention, a pre-configured digital economy transaction monitoring engine is deployed on a digital economy transaction platform. This engine, comprised of multiple modules, is capable of real-time monitoring and collection of various transaction data from the platform. The engine is configured to collect multimodal data, including but not limited to the following data types: transaction data (e.g., order amount, payment method), user behavior data (e.g., browsing history, click events), temporal data (e.g., transaction time, response time), geographic data (e.g., transaction location, IP address), and social data (e.g., user comments, interaction frequency). Once the monitoring engine begins operation, it continuously collects multimodal data on all transaction activities on the platform. The data collection process includes real-time capture of user operations on the platform, detailed information on transaction execution, and auxiliary information related to the transactions. The collected multimodal data is pre-processed to check for and remove duplicate transaction data recorded during the collection process. For example, if the same transaction is recorded multiple times due to network delays or errors, missing key fields in the records are filled or deleted. Missing data is corrected using techniques such as data interpolation and mean filling, or if excessive data is missing, relevant records are selectively deleted. Statistical methods or machine learning algorithms are used to identify and address outliers, such as unusually large transaction amounts or unusual transaction frequencies. These data points may be caused by data collection errors, system failures, or potential fraud. Standardize the cleaned digital transaction multimodal data to ensure that all data dimensions are on the same scale and eliminate differences between different data types. Standardize all numerical features (such as transaction amount and transaction frequency). Standardize all numerical features (such as transaction amount and transaction frequency). Common methods include Z-score standardization (converting data to zero mean and unit standard deviation) or Min-Max standardization (scaling data to the range of [0,1]), converting categorical features (such as payment method and product category) into numerical form, using methods such as One-Hot encoding, Label encoding, or target encoding to enable categorical data to be processed by machine learning algorithms, and standardizing timestamp and date data, such as converting time to seconds or days from a certain reference point, or extracting periodic features of time (such as hours and days of the week). The standardized digital transaction multimodal data is stored in a standardized dataset area to obtain digital transaction multimodal data.

[0058] Preferably, step S2 includes the following steps:

[0059] Step S21: performing digital transaction frequency feature analysis based on the digital transaction behavior data to generate digital transaction frequency feature data;

[0060] Step S22: performing digital transaction time series feature analysis based on the digital transaction multimodal data to generate digital transaction time series feature data;

[0061] Step S23: Designing digital transaction attention weights based on the digital transaction frequency feature data and the digital transaction time series feature data to obtain digital transaction attention weight information;

[0062] Step S24: performing attention weighting processing on the digital transaction feature data according to the digital transaction attention weight information to generate weighted digital transaction feature data.

[0063] By analyzing digital transaction behavior data for frequency characteristics, the present invention can identify frequency patterns in transaction behavior, such as the high or low frequency of certain transaction behaviors. This analysis helps improve the sensitivity of detecting abnormal behavior, for example, identifying suspicious transactions with significantly different frequencies from normal transaction behavior, thereby improving the accuracy of identifying potential abnormal behavior. Time series feature analysis based on multimodal digital transaction data can capture time series patterns in transaction behavior and help identify temporal dependencies and trends in transaction activities, such as the concentration of abnormal transactions in specific time periods. Time series feature analysis allows for more effective detection of time-dependent abnormal behavior. Combining frequency and time series feature data to design attention weights can assign appropriate weights to different transaction features, highlighting key features, particularly those that exhibit abnormal behavior in frequency and time dimensions, thereby improving the effectiveness of subsequent anomaly detection models. Weighting digital transaction feature data based on the designed attention weights can optimize the representation of feature data, allowing the model to focus more on features important for anomaly detection. For example, computational resources for anomaly detection can be allocated to new transaction data or digital transaction data corresponding to unconventional transaction behavior, thereby improving the accuracy and robustness of digital transaction anomaly detection.

[0064] In an embodiment of the present invention, a statistical analysis method is used to calculate the transaction frequency characteristics of users or user groups in digital transaction behavior data, such as the average daily number of transactions, weekly transaction fluctuations, and transaction peaks in specific time periods. The analysis also includes identifying abnormal transaction frequencies, such as a sudden increase in transaction frequency or a large number of low-amount transactions in a short period of time. The generated frequency characteristic data includes transaction frequency indicators for each user or user group, which will serve as the basic data for subsequent steps. Multimodal data is sorted and time-related features are extracted, such as the timestamp of the transaction, the duration of the transaction, the transaction interval, etc. These time series data can reveal the time distribution characteristics of transaction behavior and help identify the transaction patterns of users in different time periods. Time series analysis methods such as moving average, time series decomposition (Seasonal Decomposition), autoregressive (AR) model, etc. are used to analyze the time series characteristics of transaction behavior. The focus is on identifying transaction behavior cyclicalities (such as daily or weekly peak trading times), trend changes (such as increases or decreases in transaction frequency over time), and possible unexpected anomalies (such as large transactions at unusual times). The analysis also includes statistical analysis of transaction time intervals to identify correlation patterns between transaction frequency and time intervals. The resulting time series feature data will describe the temporal distribution patterns of each user or transaction group. Deep learning models based on attention mechanisms, such as self-attention or the Transformer model, are selected to design attention weights for digital transactions. By jointly analyzing frequency and time series feature data, the features that contribute most to transaction behavior prediction and anomaly detection are automatically identified. The frequency and time series feature data are input into the pre-designed attention mechanism model, which dynamically assigns weights based on the importance of the input features. Feature importance is automatically learned by the model based on its performance in the training data. Weight information includes the weight value of each feature, reflecting its relative importance in the overall transaction behavior analysis. For example, high-frequency trading periods for a specific user are assigned a higher weight, while regular weekday trading frequency is assigned a lower weight. This generated attention weight information guides the subsequent weighting of feature data. Based on this attention weight information, digital transaction feature data is weighted. Each feature is weighted according to its corresponding weight value. The specific steps of weighting include identifying the corresponding weight value for each feature and generating weighted feature data. This weighted data can better reflect the key features of actual trading behavior, highlight the most important features for anomaly detection, and generate weighted digital transaction feature data. This dataset contains transaction features with adjusted weights, which can effectively improve the model's performance in subsequent prediction and detection.

[0065] Preferably, step S3 includes the following steps:

[0066] Step S31: Obtain historical digital trading strategies;

[0067] Step S32: defining a digital transaction strategy state space based on historical digital transaction strategies to obtain a digital transaction strategy state space;

[0068] Step S33: Analyzing the transaction strategy correlation characteristics of the historical digital transaction strategies to obtain transaction strategy correlation characteristic data;

[0069] Step S34: defining the digital trading strategy action space according to the trading strategy associated feature data and the digital trading strategy state space to obtain the digital trading strategy action space;

[0070] Step S35: Designing a digital transaction strategy network according to the digital transaction strategy action space to obtain a digital transaction strategy network;

[0071] Step S36: Designing a digital transaction value network based on the digital transaction strategy network and historical digital transaction strategies to obtain a digital transaction value network;

[0072] Step S37: Designing a reinforcement learning mathematical model for the digital transaction strategy based on the digital transaction strategy network and the digital transaction value network to establish a digital transaction strategy reinforcement learning model;

[0073] Step S38: transmitting the weighted digital transaction feature data to the digital transaction strategy reinforcement learning model for digital transaction strategy prediction processing to generate digital transaction strategy prediction data.

[0074] The present invention obtains historical digital trading strategies, leveraging past trading data and accumulated experience to provide a reliable foundation for model construction. The collection and analysis of historical data is used to identify previously effective strategies, which in turn provides a reference for current trading processing and anomaly detection, ensuring a certain historical basis when processing new data, helping to improve the accuracy and robustness of predictions. By defining the state space of digital trading strategies, all trading states and their relationships can be clearly depicted. State space definition is a core step in reinforcement learning. It helps the subsequently established reinforcement learning model clarify the various states of the current trading environment and associate different strategy options with each state. This ensures that the model can effectively identify and adapt to different trading states when faced with complex trading environments, thereby optimizing the detection and decision-making process. By analyzing the correlation characteristics of historical digital trading strategies, the inherent connections between different strategies can be discovered, key features can be identified, and factors that influence trading behavior prediction can be extracted and summarized. This provides data support for the subsequent definition of the action space and significantly improves the pertinence and effectiveness of strategy selection. By defining the action space of a strategy, it is possible to clearly define the strategies or actions that can be taken under different states. This definition of the action space enables the model to evaluate multiple strategy options in each trading state and optimize their execution performance, addressing various complex trading scenarios and potential anomalies, and improving overall detection effectiveness. By designing a digital trading strategy network, it is possible to systematically represent and process the complex relationships in the strategy space. The establishment of the strategy network enables the system to efficiently select and optimize strategies. The network structure captures the nonlinear relationships and interactions between strategies, improving not only the efficiency of policy reasoning but also the intelligent decision-making capabilities of the subsequent reinforcement learning model. By designing a value network, the long-term performance (i.e., value) of different strategies is evaluated and quantified. The construction of the value network provides an evaluation metric for the model, enabling the system to select strategies that optimize long-term returns when faced with multiple strategies. This improves the decision-making quality of the subsequent reinforcement learning model in trading strategy prediction, allowing anomaly detection to focus not only on short-term but also long-term effects, ensuring the reliability and sustainability of detection results. By combining the policy network and the value network to design a reinforcement learning model, the system can adaptively optimize trading strategies, predicting subsequent trading strategies based on trading data. The reinforcement learning model optimizes strategies through continuous experimentation and feedback, ensuring that the model can continuously learn and improve in dynamic and complex trading environments, thereby improving detection accuracy and efficiency and reducing false positives and missed detections. By inputting weighted feature data into the reinforcement learning model, the system can use the optimized strategy to make efficient predictions. The data generated by the prediction process provides a precise source of information for subsequent anomaly detection and strategy adjustments, ensuring that the detection model can respond in real time and accurately, improving the system's performance in complex trading environments and providing strong support for final anomaly detection.

[0075] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed implementation steps of step S3 are shown in the flowchart. In this embodiment, step S3 includes:

[0076] Step S31: Obtain historical digital trading strategies;

[0077] In this embodiment, historical digital transaction strategy data is extracted from the transaction platform's historical records. This strategy data includes transaction behavior decisions taken by users in different time periods and transaction scenarios. For example, transaction strategies derived from different combinations of transaction behaviors in historical digital transactions include key information such as transaction time, amount, participants, payment method, and geographical distribution, as well as how these factors affect the ultimate transaction success rate and user satisfaction. This historical transaction strategy data is archived to form a structured database to obtain historical digital transaction strategies.

[0078] Step S32: defining a digital transaction strategy state space based on historical digital transaction strategies to obtain a digital transaction strategy state space;

[0079] In an embodiment of the present invention, historical digital transaction strategies are analyzed to determine the main factors affecting the selection of transaction strategies. When defining the state space, these factors are used as state variables. These variables include transaction amount ranges, user historical behavior characteristics, time characteristics, geographic location, etc. The state space is composed of a combination of these variables. To simplify processing, the continuous state variables are discretized, and each state in the state space is encoded and stored to obtain the digital transaction strategy state space.

[0080] Step S33: Analyzing the transaction strategy correlation characteristics of the historical digital transaction strategies to obtain transaction strategy correlation characteristic data;

[0081] In this embodiment of the present invention, correlation feature analysis is performed on historical digital trading strategies under each state to identify which trading strategies have strong correlations with specific state variables. For example, the correlation between each trading state and the trading strategy in the historical trading strategy is analyzed, and association rule analysis (such as the Apriori algorithm) or correlation analysis techniques are used to extract feature data with significant correlations. Trading strategy correlation feature data is generated, including information about the trading strategies associated with each trading state. This feature data is aggregated to form the basis for subsequent action space definition.

[0082] Step S34: defining the digital trading strategy action space according to the trading strategy associated feature data and the digital trading strategy state space to obtain the digital trading strategy action space;

[0083] In this embodiment of the present invention, after defining the state space and associated features, the action space of the digital trading strategy is defined. The action space includes all trading strategy options available in a specific state, that is, the trading action behavior paths for a specific trading state in the digital trading strategy state space. The action space is stored in a matrix or table format, with each row representing a state and each column representing an action within that state. The stored information includes the historical execution results of each action, thereby obtaining the digital trading strategy action space.

[0084] Step S35: Designing a digital transaction strategy network according to the digital transaction strategy action space to obtain a digital transaction strategy network;

[0085] In an embodiment of the present invention, a digital trading policy network (Policy Network) is designed based on the action space. The network architecture adopts a deep neural network. The input layer receives state space data, the hidden layer models the complex relationship between state and action, and the output layer generates corresponding trading policy action recommendations. The network design takes into account the nonlinear relationship between different states and the interaction between high-dimensional features. Efficient policy reasoning is achieved through structures such as multi-layer perceptrons (MLP) or convolutional neural networks (CNN). The policy network is trained using state-action pairs in historical data. Through supervised learning, the network gradually optimizes its parameters so that it can output the optimal trading policy recommendations under a given state. During the training process, cross-validation technology is used to avoid overfitting and ensure the good performance of the network on unknown data to obtain a digital trading policy network.

[0086] Step S36: Designing a digital transaction value network based on the digital transaction strategy network and historical digital transaction strategies to obtain a digital transaction value network;

[0087] In an embodiment of the present invention, a digital transaction value network (Value Network) is designed to evaluate the long-term effects of each transaction strategy. The Value Network inputs state and action data and outputs a value score for each strategy, that is, the expected benefit or effect of the strategy in long-term execution. The network structure is similar to the digital transaction strategy network, but the output layer has only one node, which represents the expected value under a given state and strategy. The Value Network is trained using historical strategy data, with the goal of predicting the true value of each strategy under a given state. During the training process, the error between the predicted value and the actual value is minimized by adjusting the network weights (using the mean square error as the loss function). The network training adopts an iterative optimization method (such as gradient descent) and uses regularization to prevent overfitting, ensuring the stability and generalization ability of the model, so as to obtain a digital transaction value network.

[0088] Step S37: Designing a reinforcement learning mathematical model for the digital transaction strategy based on the digital transaction strategy network and the digital transaction value network to establish a digital transaction strategy reinforcement learning model;

[0089] In an embodiment of the present invention, a digital trading strategy model based on reinforcement learning is established by combining the policy network and the value network, using deep Q learning (DQN) or the policy gradient method (Policy Gradient) as the core algorithm. The goal of the reinforcement learning model is to optimize the digital trading strategy through continuous experimentation and feedback. The input of the model is the state space feature, and the output is the optimal strategy action. The model continuously updates the parameters of the policy network and the value network by interacting with the environment, so that the strategy with the greatest long-term benefit can be selected under different states. The reinforcement learning model is trained in a simulation environment, which generates different trading scenarios based on historical trading data. The model tests different strategy actions in these scenarios and updates the network parameters based on the feedback obtained (such as user reactions, transaction success rate, etc.). The model gradually optimizes its strategy through multiple rounds of iteration, so that it can make the best decision when facing a complex and changing trading environment.

[0090] Step S38: transmitting the weighted digital transaction feature data to the digital transaction strategy reinforcement learning model for digital transaction strategy prediction processing to generate digital transaction strategy prediction data.

[0091] In this embodiment of the present invention, weighted digital transaction feature data is input into a trained reinforcement learning model. The model calculates based on the input feature data and infers the optimal strategic action for the current transaction state. The reinforcement learning model generates prediction data by comprehensively considering factors such as current feature information, historical strategy effectiveness, and long-term value. The model outputs digital transaction strategy prediction data, which contains recommendations for the optimal strategy under the current transaction environment. This generated strategy prediction data is then fed into the next step or directly applied to transaction decision-making, thereby improving transaction success rates and user experience.

[0092] Preferably, step S36 includes the following steps:

[0093] Step S361: dividing the historical digital transaction strategies into historical transaction strategy action nodes according to the digital transaction strategy network to obtain historical transaction strategy action nodes;

[0094] Step S362: performing action node status evaluation on the historical transaction strategy action node to obtain historical transaction strategy action node status evaluation data;

[0095] Step S363: Design a digital transaction value network based on historical transaction strategy action node status evaluation data to obtain a digital transaction value network.

[0096] The present invention divides historical digital trading strategies into action nodes, which can decompose complex trading strategies into smaller and more manageable action nodes. Each action node represents a specific trading decision or behavior unit. This division method helps to refine and clarify the components of different trading strategies, more accurately identify key behavior nodes in the trading process, and improve the sensitivity and accuracy of anomaly detection. A state evaluation is performed on each action node to quantify the performance and role of each node in different historical trading strategies. The state evaluation gives each node a clear indicator based on the transaction status in the historical trading strategy, reflecting its effectiveness and risk in the strategy execution process, identifying which nodes are high-risk and which nodes are crucial to the successful detection and prevention of abnormal behavior, and providing key data support for the subsequent design of the value network. By using the state evaluation data of the action nodes to design the value network, the system can systematically evaluate the long-term effects of different trading strategies. The design of the value network not only takes into account the immediate benefits of each action node, but also takes into account its long-term impact. The establishment of the value network ensures that the system can quantify the transaction behavior status of each predicted transaction decision in a dynamic and complex transaction environment, so as to predict the subsequent digital transaction decisions of each weighted digital transaction feature data, so as to improve the ability to detect predictive anomalies.

[0097] In an embodiment of the present invention, a previously constructed digital trading strategy network is utilized to analyze each decision point in historical digital trading strategies. Each decision point corresponds to a specific trading behavior or strategy choice. By analyzing these decision points, each historical strategy is divided into multiple action nodes. Each node represents a specific decision or operation. These operations are connected through different paths and nodes in the trading strategy network, reflecting the execution order and logical relationships of the strategy. Within each strategy execution path, the strategy is split into multiple consecutive action nodes. A similar division operation is performed on all historical strategies to generate a series of historical trading strategy action nodes. These nodes represent all possible decision points that occurred during the trading process. An action node status evaluation model is constructed to assess the performance and effectiveness of each action node in the historical trading strategy. The evaluation model uses machine learning algorithms, such as random forests or support vector machines, to analyze the status performance of different nodes. Evaluation metrics include each node's execution frequency, success rate, execution time, and user response. These metrics can reflect the role and importance of each action node in the trading process. Historical transaction decision data is fed into the evaluation model, which performs a detailed status evaluation of each action node. The evaluation model analyzes historical data to calculate key status data for each node, such as the node's performance under specific circumstances (e.g., peak periods or promotional events). The evaluation model outputs status evaluation data for each node, categorized as "efficient," "inefficient," or "critical," helping to identify key action nodes within the trading strategy. All evaluation results are aggregated to generate historical transaction strategy action node status evaluation data. This data is stored in a table or matrix format, containing the status evaluation results for each action node. This data is used in subsequent value network design to ensure that the value network accurately reflects the actual performance of the trading strategy. Based on the action node status evaluation data generated in the previous step, a digital transaction value network is designed. The goal of the value network is to evaluate and predict the overall performance (i.e., expected value) of different strategies over the long term, including their impact on transaction success rate, user satisfaction, and long-term user retention. A deep neural network architecture is designed as the foundation of the digital transaction value network. The network's input layer receives action node status evaluation data, including characteristics such as each node's status score and execution performance. The hidden layer nonlinearly combines these input data to uncover deep connections between different nodes. The output layer provides an overall value score for each trading strategy, reflecting the strategy's expected long-term returns under different states. The network uses architectures such as multi-layer perceptrons (MLPs) or recurrent neural networks (RNNs) to assess the value of complex strategies. The value network is trained using historical trading data, with the goal of minimizing the error between the network's predictions and the actual strategy results. Through multiple iterations, the network's weights and biases are adjusted, enabling the network to accurately predict the long-term value of different strategies.Cross-validation and regularization techniques are used during network training to prevent overfitting and ensure good generalization when processing new data. After training, the generated digital transaction value network will be used to evaluate and predict the performance of different strategies in digital transactions in real time. This will be applied to reinforcement learning models to predict the selection of trading strategies for real-time digital transactions.

[0098] Preferably, step S4 includes the following steps:

[0099] Step S41: performing multi-dimensional digital transaction combination analysis processing based on the digital transaction strategy prediction data to obtain multi-dimensional digital transaction combination data;

[0100] Step S42: performing a multi-dimensional digital transaction fraud feature analysis based on the multi-dimensional digital transaction combination data to generate multi-dimensional digital transaction fraud feature data;

[0101] Step S43: performing fraud feature derivation processing on the multi-dimensional digital transaction fraud feature data using a preset generative adversarial network model to generate multi-dimensional digital transaction fraud feature derived data;

[0102] Step S44: Designing a mathematical model for the adversarial game of abnormal transactions based on the multi-dimensional digital transaction fraud feature-derived data to generate an abnormal transaction adversarial game model.

[0103] The present invention performs multi-dimensional combined analysis on digital transaction strategy prediction data and comprehensively analyzes transaction behavior in different dimensions. It captures the complex and diverse interaction patterns in transaction behavior and reveals abnormal features that cannot be discovered by single-dimensional analysis. It performs fraud feature analysis on multi-dimensional digital transaction combination data, analyzes potential fraud features in transaction behavior, and identifies abnormal behaviors hidden in normal transaction patterns, providing an important basis for further anomaly detection, enhancing the system's detection capabilities when dealing with complex and hidden fraud behaviors, and effectively reducing false alarm rates and missed reports. By using a generative adversarial network (GAN) to derive multi-dimensional fraud feature data, the system can simulate potential complex fraud behaviors and generate new fraud features. This derivative processing helps the system introduce more diverse fraud patterns in model training, improve the robustness and generalization ability of the model, better predict and prevent new and variant fraud behaviors, and ensure that the detection model always maintains high accuracy and adaptability. Based on multi-dimensional fraud feature derivative data, an adversarial game model is designed to simulate and predict the game process between fraudsters and the detection system. The adversarial game model helps the system develop more effective detection strategies by analyzing the evasion strategies that fraudsters may adopt, improving the system's automated detection capabilities for complex and abnormal transactions. It also effectively responds to the dynamic changes and evolution of fraudulent behavior, ensuring that the system can continue to optimize and maintain its advantage in a fierce adversarial environment.

[0104] In an embodiment of the present invention, relevant data is extracted from digital transaction strategy prediction data. This data includes predicted transaction strategies and corresponding feature data, such as user behavior characteristics and transaction environment characteristics. Multidimensional data analysis techniques are then used to perform a combined analysis of the extracted data. Combined analysis involves analyzing the various combinations of digital transaction strategies sent within the digital transaction strategy prediction data and identifying interactions between different dimensions, as well as the impact of these interactions on transaction behavior and strategy selection. The analysis results in multidimensional digital transaction combination data, which reflects the complex connections and potential patterns between dimensions. This multidimensional digital transaction combination data is then analyzed to identify fraud signatures. For example, this analysis can identify whether there are frequent small transactions within a specific time period, or large fund transfers between multiple geographic locations within a short period of time. Feature analysis not only focuses on anomalies in a single dimension, but also considers how interactions between multiple dimensions lead to the occurrence of abnormal transaction behavior. The resulting fraud signature data reflects these complex patterns and potential risks. The analysis results are integrated to generate multidimensional digital transaction fraud signature data, which includes fraudulent behavior for each transaction. A preset generative adversarial network (GAN) model is constructed, which includes two core parts: a generator and a discriminator. The generator is responsible for generating possible fraud feature data, while the discriminator is responsible for distinguishing these generated data from real fraud feature data. Through multiple rounds of adversarial training, the generator gradually learns to generate fraud features similar to real data, and the discriminator continuously improves its discrimination ability, so that the entire network is continuously optimized in the generation and detection of fraud features. The multi-dimensional digital transaction fraud feature data is input into the preset GAN model. The generator generates new fraud feature data based on these input data. The generated data contains potential fraud patterns that have not been detected in the original data, helping the system to identify more covert or new fraud behaviors. The generated multi-dimensional digital transaction fraud feature data contains the original feature data and the new fraud features generated by the GAN model. When designing the adversarial game model, we first established a theoretical foundation based on fundamental game theory concepts such as Nash equilibrium and evolutionarily stable strategies (ESS). The model aims to simulate the dynamic confrontation between fraudsters and defense systems, helping to predict the evasion strategies employed by fraudsters. By leveraging data derived from multi-dimensional digital transaction fraud characteristics, we construct an adversarial game model for abnormal transactions. The model involves two participants: the fraudster and the detection system. The fraudster's goal is to maximize their transaction success rate, while the detection system's goal is to maximize its detection accuracy. Through multiple rounds of iteration, the model simulates the performance of both parties under different strategies and analyzes the possible strategy combinations they may adopt during the game. The results of each round are fed back into the model to adjust the strategic choices of both parties.The adversarial game model is trained using historical data and generated derivative data, enabling it to accurately predict game outcomes under different conditions. During the model optimization process, reinforcement learning algorithms are used to further enhance the model's decision-making capabilities in complex game environments. Through continuous optimization, the model can capture more complex adversarial strategies adopted by fraudsters and adjust the detection system's strategy to meet these challenges. After training is completed, the final abnormal transaction adversarial game model is generated, providing the system with accurate fraud response strategies to effectively respond to and detect complex fraudulent behaviors in actual trading environments.

[0105] Preferably, step S5 includes the following steps:

[0106] Step S51: performing abnormal transaction adversarial game feature analysis and processing according to the abnormal transaction adversarial game model to generate abnormal transaction adversarial game feature data;

[0107] Step S52: Designing digital transaction fraud tracing tree nodes based on the abnormal transaction adversarial game feature data to obtain digital transaction fraud tracing tree node data;

[0108] Step S53: Designing associated labels for a fraud tracing tree structure based on the weighted digital transaction feature data to generate associated label data for the fraud tracing tree structure;

[0109] Step S54: setting fraud tracing tree structure weight parameters according to the digital transaction attention weight information corresponding to the weighted digital transaction feature data to obtain the fraud tracing tree structure weight parameters;

[0110] Step S55: Designing a step-by-step fraud tracing tree structure model for digital transactions based on the fraud tracing tree structure associated tag data for the digital transaction fraud tracing tree node data to establish a preliminary digital transaction fraud tracing tree model;

[0111] Step S56: Optimizing the tree structure weight parameters of the preliminary digital transaction fraud tracing tree model according to the fraud tracing tree structure weight parameters to obtain the digital transaction fraud tracing tree model.

[0112] The present invention uses feature analysis based on the abnormal transaction confrontation game model. The system can deeply explore the strategic confrontation relationship between the fraudster and the detection system, extract key feature data reflecting the confrontation process between the two parties, and provide an important reference basis for subsequent detection and response strategies. The generated confrontation game feature data can not only improve the accuracy of abnormal detection, but also help the system predict the trend of changes in fraudulent behavior, thereby improving the overall prevention capability. Designing the nodes of the fraud tracing tree based on the confrontation game feature data can effectively decompose complex abnormal transaction behaviors into multiple key nodes, each node representing a specific fraud behavior feature or transaction pattern, so that the tracing tree can display the entire fraud behavior chain in a modular manner, which helps the system to quickly locate and identify the source of abnormal behavior and improve the efficiency of tracing and analysis. The fraud tracing tree structure is designed with associated labels. The system can assign appropriate labels to each node and branch to clarify its role and relationship in the entire tracing structure, closely associate different transaction features with specific fraud behavior patterns, thereby enhancing the readability and interpretability of the tracing tree when identifying complex fraud behaviors, and ensuring that the system can obtain a clearer and more comprehensive perspective when conducting tracing analysis. By combining the attention weight information of weighted digital transaction feature data to set the weight parameters of the traceability tree, the system can assign appropriate weights to different nodes and paths. The setting of weight parameters ensures that the system can more accurately focus on key features and important paths during traceability analysis, thereby improving the accuracy and efficiency of the overall model in tracing fraud behavior, and enhancing the adaptability and decision-making ability of the traceability tree in complex transaction environments. By gradually designing and constructing a fraud tracing tree model, it is possible to refine and expand the structure of the traceability tree layer by layer, starting from the basic nodes, to ensure that the model can adapt to the needs of tracing fraud behaviors of different scales and complexities, and can flexibly adjust the distribution of nodes and paths to adapt to changing transaction patterns and fraud methods. The preliminary fraud tracing tree model provides the system with a dynamic and flexible tool that can be gradually evolved and improved in practical applications. By optimizing the weight parameters of the initially designed traceability tree model, the system can further improve the accuracy and reliability of the model. The optimized tree structure weight parameters ensure that the system can more effectively identify key nodes of fraudulent behavior and give priority to paths with higher risks, such as focusing on higher transaction frequencies or recent digital transactions. The weight optimization process makes the traceability tree model perform better in actual applications and enhances the system's prevention capabilities and response speed in high-risk transaction environments.

[0113] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 Detailed implementation steps of step S5 are shown in the flowchart. In this embodiment, step S5 includes:

[0114] Step S51: performing abnormal transaction adversarial game feature analysis and processing according to the abnormal transaction adversarial game model to generate abnormal transaction adversarial game feature data;

[0115] In this embodiment of the present invention, an abnormal transaction adversarial game model is used to conduct in-depth analysis of historical transaction data and derived fraud feature data. By simulating the strategic confrontation between fraudsters and defense systems, adversarial features of various transaction behaviors are extracted. This includes identifying abnormal behavior characteristics under different strategy combinations, such as the repeated use of a certain transaction pattern by fraudsters in specific environments, or how fraudsters adjust their strategies to evade detection. Through analysis of the adversarial game model, abnormal transaction adversarial game feature data is generated, including detailed adversarial features such as frequent changes in transaction paths and sudden large-value transactions. These features reflect the evolution of fraudster strategies during multiple confrontations, ensuring that subsequent retrospective analysis can accurately capture key abnormal behaviors.

[0116] Step S52: Designing digital transaction fraud tracing tree nodes based on the abnormal transaction adversarial game feature data to obtain digital transaction fraud tracing tree node data;

[0117] In this embodiment of the present invention, nodes of a traceability tree are defined based on the characteristic data of abnormal transaction adversarial game. Each node represents a specific abnormal behavior or transaction feature. These nodes are connected by paths to form a complete chain of fraudulent behavior. The design takes into account the causal relationship between each node and the strategy switching during the game process. All extracted adversarial game features are classified and organized to generate node data for the digital transaction fraud traceability tree. This data includes a description of each node's characteristics, associated transaction behavior, and its correlation with other nodes. The node data will be used in the structural design of the traceability tree to ensure that each node accurately reflects the fraudulent behavior.

[0118] Step S53: Designing associated labels for a fraud tracing tree structure based on the weighted digital transaction feature data to generate associated label data for the fraud tracing tree structure;

[0119] In this embodiment of the present invention, weighted digital transaction feature data is used to design labels for each node in the traceability tree. The labels serve to correlate each transaction action combination with the predicted transaction action, helping the system quickly identify key fraudulent action nodes in digital transactions during the traceability process. Label design is based on the correspondence between each digital transaction combination in the weighted digital transaction feature data, mapping the feature data to the nodes of the traceability tree. The designed labels are attached to each node in the traceability tree to generate associated label data for the fraud traceability tree structure. This associated label data is then used to associate the nodes in the traceability tree.

[0120] Step S54: setting fraud tracing tree structure weight parameters according to the digital transaction attention weight information corresponding to the weighted digital transaction feature data to obtain the fraud tracing tree structure weight parameters;

[0121] In an embodiment of the present invention, the weight parameters of each node and path in the traceability tree are set based on the weighted digital transaction feature data and the corresponding attention weight information. The weight parameters determine which paths or nodes the system should prioritize during the traceability process. The attention weight information of each feature is taken into account during the setting process, so that the nodes corresponding to high-weight features have a higher priority in the traceability tree. In this way, the system can quickly lock on to key paths during the traceability process, such as key paths in high-frequency transactions or recent transactions, thereby improving efficiency. The set weight parameters are saved as the core attributes of the traceability tree structure. These parameters will be used in the preliminary design and optimization of the traceability tree model to ensure that the model can accurately reflect the real transaction risks.

[0122] Step S55: Designing a step-by-step fraud tracing tree structure model for digital transactions based on the fraud tracing tree structure associated tag data for the digital transaction fraud tracing tree node data to establish a preliminary digital transaction fraud tracing tree model;

[0123] In an embodiment of the present invention, a step-by-step fraud tracing tree structure model for digital transactions is designed based on the fraud tracing tree structure-associated label data for the digital transaction fraud tracing tree node data. During the design process, each node is connected according to causality and correlation to form a complete tracing path tree structure. The structural model design takes into account multiple factors, including the direct association between nodes, the influence of weight parameters on path priority, and the guiding role of label data on model branches. When designing the preliminary model, a step-by-step construction and verification method is adopted to construct a simplified tracing tree structure, conduct preliminary fraud tracing analysis, and gradually expand and optimize the model based on the analysis results. The model expansion at each stage is verified by actual transaction data to ensure that the model has sufficient accuracy and breadth when identifying fraudulent behavior, so as to establish a preliminary digital transaction fraud tracing tree model.

[0124] Step S56: Optimizing the tree structure weight parameters of the preliminary digital transaction fraud tracing tree model according to the fraud tracing tree structure weight parameters to obtain the digital transaction fraud tracing tree model.

[0125] In an embodiment of the present invention, global weight parameters are optimized based on a preliminarily designed tracing tree model. The optimization goal is to adjust the weights of each node and path so that the model can prioritize high-risk paths during the tracing process and reduce the waste of resources on low-risk paths. An optimization algorithm (such as a genetic algorithm or a particle swarm optimization algorithm) is used to adjust the weight parameters. The best weight allocation is found by simulating multiple transaction scenarios. After the optimization is completed, the final digital transaction fraud tracing tree model is obtained. This model has an optimized weight parameter setting and can efficiently and accurately identify and trace complex fraudulent behaviors in practical applications.

[0126] Preferably, step S51 includes the following steps:

[0127] Step S511: performing abnormal transaction game evolution processing according to the abnormal transaction adversarial game model to generate abnormal transaction game evolution data;

[0128] Step S512: performing game evolution replication dynamic processing according to the abnormal transaction game evolution data to generate abnormal transaction game evolution replication dynamic data;

[0129] Step S513: performing game evolution equilibrium point analysis and processing based on the abnormal transaction game evolution replication dynamic data to generate game evolution equilibrium point data;

[0130] Step S514: performing abnormal transaction adversarial game feature analysis on the abnormal transaction adversarial game model according to the game evolution equilibrium point data to generate abnormal transaction adversarial game feature data.

[0131] The present invention applies game evolution processing to abnormal transactions, simulating and tracking the strategic evolution between fraudsters and detection systems. This reveals the changing trends in the strategies of both parties in the ongoing confrontation, helping to identify the evolutionary paths and emerging patterns of fraudulent behavior. This allows anomaly detection to move beyond static rules and adapt in real time to changes in fraudulent strategies, thereby enhancing the system's foresight and early warning capabilities. Through dynamic replication of game evolution, the spread and reinforcement effects of different strategies during the game are quantified and evaluated, understanding why certain fraudulent strategies are widely adopted or abandoned in specific environments. The generated dynamic replication data provides a quantitative basis for further analysis of strategy effectiveness and risk, prioritizing those strategies that have evolved into mainstream fraudulent methods, thereby improving proactive and targeted defenses. Through equilibrium point analysis, the stable strategy combinations that the two parties in the game can achieve in a long-term confrontation, namely the Nash equilibrium, can be determined. This equilibrium point data reveals the equilibrium state of the optimal strategies of the fraudster and the detection system under specific conditions. This allows the system to predict the fraudster's behavior patterns and optimize its own detection strategy, ensuring that the system maintains the initiative in complex transaction environments and avoids being passive in strategic confrontations. Through game feature analysis based on equilibrium point data, we can deeply explore the key features of fraudulent behavior and optimize the anomaly detection model accordingly. The generated game feature data reflects which features are most critical for identifying and preventing fraudulent behavior during the game process, greatly improving the accuracy and efficiency of anomaly detection. When faced with changing fraud methods, the system can accurately identify core risk points, respond quickly, and enhance overall defense capabilities.

[0132] In an embodiment of the present invention, based on the abnormal transaction adversarial game model, the strategy sets of all parties involved (such as fraudsters and detection systems) are initialized. Each party has a set of strategies, and these strategies will continue to evolve during the game. The initial strategies include common fraud behaviors (such as small-amount multiple transactions, rapid cross-regional fund transfers) and corresponding detection strategies (such as transaction frequency analysis, cross-border transaction monitoring, etc.). By simulating multiple rounds of games, the performance of participants under different strategy combinations is recorded. After each round of games, both parties adjust their own strategies based on the other party's strategies, thereby promoting the evolution of strategies. The replicator dynamics equation in evolutionary game theory is used to describe the changing trends of strategies. The replicator dynamics processing aims to simulate the propagation and changes of strategies during the game. This equation can predict which strategies will be retained and strengthened, and which strategies will gradually be eliminated. During the evolutionary process, the results of each round of the game are recorded, including key information such as strategy selection, success rate, and cost. During each round of dynamic replication, key information such as changes in strategy proportions and increases or decreases in success rates are recorded to generate dynamic replication data for the abnormal trading game evolution. This dynamic data reflects how strategies evolve and adapt to changes in the external environment during the game, providing a foundation for subsequent equilibrium analysis. Based on the replicated dynamic data, a model for analyzing the game's evolutionary equilibrium is constructed. This model uses game theory tools such as Nash equilibrium analysis and evolutionarily stable strategies (ESS) to analyze the stability of different strategy combinations. The model aims to identify equilibrium states that can be reached by both strategies during the game, i.e., states where neither party has the incentive to unilaterally change their strategy. By simulating multiple games, the equilibrium point is gradually approached. The strategy distribution at the end of each game is analyzed to determine when the system reaches a stable state. For each possible strategy combination, its stability in the equilibrium state is calculated. Key characteristics of each equilibrium point, such as the strategy combination, expected return, and risk exposure, are recorded and stored to generate game evolutionary equilibrium data. Using data from the game's evolving equilibrium points, we analyze the characteristics and impact of each strategy at equilibrium. This feature analysis focuses on how strategies evolve to equilibrium during the game and the correlations between different strategies at equilibrium. By comparing data from different equilibrium points, we identify high-risk strategy combinations, which often represent effective strategies that fraudsters are likely to use repeatedly. The analysis results are converted into characteristic data for the abnormal transaction countermeasures game. This data includes fraudulent behavior at each equilibrium point, the strategy combinations for addressing fraudulent behavior, and the key nodes that trigger abnormal behavior.

[0133] Preferably, step S6 includes the following steps:

[0134] Step S61: transmitting the weighted digital transaction feature data to the digital transaction fraud tracing tree model for digital transaction anomaly prediction processing to generate digital transaction anomaly prediction data;

[0135] Step S62: performing digital transaction anomaly assessment processing on the digital transaction anomaly prediction data based on a preset digital transaction anomaly assessment decision to generate digital transaction anomaly assessment data;

[0136] Step S63: Mapping the digital transaction anomaly assessment data to the digital transaction anomaly prediction data to perform digital transaction anomaly feedback information integration processing to generate digital transaction anomaly feedback data;

[0137] Step S64: Transmitting the digital transaction abnormality feedback data to the terminal to execute the digital transaction abnormality feedback operation.

[0138] The present invention transmits weighted digital transaction feature data to the fraud tracing tree model, which can make full use of the feature weight information constructed previously to perform accurate abnormal transaction prediction processing, ensuring that the prediction model can be analyzed based on the most relevant feature data, thereby generating high-precision abnormal prediction results, effectively improving the accuracy and sensitivity of the prediction, and enabling the system to promptly detect and warn of potential abnormal transaction behaviors. By processing the prediction data using preset abnormality assessment decisions, the prediction results are further refined and evaluated, which helps to filter out false alarms, ensure that only alarms are issued for truly abnormal transaction behaviors, reduce the false alarm rate, and provide quantitative information on fraudulent behavior, thereby improving the detection reliability and accuracy of the overall system. Mapping and integrating the abnormality assessment data with the prediction data can form comprehensive abnormal feedback information, ensuring that the feedback data is complete and consistent, enabling the system to accurately locate abnormal behavior and provide detailed references for further preventive measures. It can not only help operators quickly understand the problem, but also provide valuable feedback information for further optimization of the model. By transmitting abnormal feedback data to the terminal, the system can trigger corresponding preventive or remedial measures in real time or in advance, quickly responding to potential risks. This ensures that the system not only detects abnormal transactions but also takes effective intervention or prevention measures, thereby protecting the security of the digital transaction environment. The real-time feedback mechanism provides the system with closed-loop control capabilities, enabling it to continuously optimize its detection and protection capabilities, forming an adaptive, self-optimizing security system.

[0139] In this embodiment of the present invention, weighted digital transaction feature data is input into a pre-built digital transaction fraud tracing tree model. Based on the fraud features and weight data generated in the previous steps, this model is capable of predicting complex transaction behaviors. During model initialization, the feature data is loaded to ensure the model fully utilizes this weight information for prediction. After receiving the input data, the tracing tree model analyzes the input feature data based on the weight parameters of each node to determine whether it conforms to known fraud patterns or exhibits abnormal behavior. Based on the interconnected paths between nodes, the model predicts abnormal transaction behavior. The prediction results include abnormal behavior features or codes for each transaction, which reflect the specific abnormality or fraudulent behavior, thereby generating digital transaction anomaly prediction data. The predicted anomaly data is further evaluated based on the system's preset digital transaction anomaly assessment decision criteria. These criteria may include common anomaly indicators such as abnormal transaction amounts, frequent transactions, and sudden changes in geographic location. Based on these preset criteria, the system develops multi-level evaluation rules to ensure accurate evaluation indicators for each anomaly. The predicted anomaly data is then compared against the evaluation criteria, and the evaluation model determines the degree of abnormality for each transaction. The system then assigns different quantitative indicators, such as risk factors, based on the different anomaly characteristics. The assessment process considers multiple factors, including historical transaction patterns, changes in user behavior, and the transaction environment, ensuring highly accurate and comprehensive results. This generates digital transaction anomaly assessment data, which records the assessment results and risk scores for each transaction. This data is then mapped and integrated with previously generated anomaly prediction data. By comparing and integrating the predictions with the assessment results, the system generates more comprehensive anomaly feedback information. This integration process combines various risk indicators from the assessment with anomaly characteristics from the predictions to generate digital transaction anomaly feedback data, including specific anomaly or fraudulent behavior, anomaly scores, and response strategies. The generated digital transaction anomaly feedback data is transmitted to a designated terminal system. This terminal may include an automated risk control system, a user notification platform, or a manual review system. This data transmission process ensures the accuracy and timeliness of the information, enabling the terminal system to respond promptly to anomaly feedback information. Within the terminal system, the received anomaly feedback data triggers the corresponding execution actions. For example, the automated system freezes suspicious accounts, blocks risky transactions, or sends warning notifications to users based on feedback data. After the execution process is completed, the terminal system will feed back the execution results (such as successful operation, user response, etc.) to the main system to improve the closed-loop control of transactions and continuously optimize the anomaly detection process.

[0140] The present invention provides a beneficial effect in that the machine learning-based digital economic transaction processing and anomaly detection method of the present invention, by introducing a machine learning model, particularly a reinforcement learning model, can adaptively adjust digital transaction strategies and detection methods. Unlike traditional methods that rely on predefined rules and thresholds, the machine learning model extracts features from historical data and predicts future transaction behavior through continuous learning and optimization. This dynamic adaptability enables the system to promptly detect and respond to emerging fraud patterns, no longer limited to past experience and known fraud patterns. By designing digital transaction attention weight information and adversarial game mathematical models, the system can handle more complex and hidden anomalies. The weighted processing of attention weights enables the system to focus resources on analyzing anomalous transaction features, improving the accuracy and efficiency of anomaly detection. The introduction of adversarial game models enables the system to simulate and counter potential fraud strategies and identify complex fraud behaviors that circumvent traditional detection methods. Furthermore, during the anomaly detection process, the model is continuously optimized. By optimizing the weight parameters of the preliminary digital transaction fraud tracing tree model, the model's accuracy and effectiveness are continuously improved. The continuous learning and optimization capabilities keep pace with evolving fraud behaviors, gradually enhancing the detection capabilities of new and complex digital transaction anomalies.

[0141] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0142] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A digital economic transaction processing and anomaly detection method based on machine learning, characterized in that: The following steps are involved: Step S1: collecting digital transaction multimodal data from a digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain digital transaction multimodal data; performing digital transaction feature analysis and processing on the digital transaction multimodal data to generate digital transaction feature data; Step S2: formulating digital transaction attention weight information based on the digital transaction multimodal data; performing attention weight weighting processing on the digital transaction feature data according to the digital transaction attention weight information to generate weighted digital transaction feature data; Step S3: obtaining historical digital transaction strategies; establishing a digital transaction strategy reinforcement learning model based on the historical digital transaction strategies; transmitting weighted digital transaction feature data to the digital transaction strategy reinforcement learning model for digital transaction strategy prediction processing to generate digital transaction strategy prediction data; wherein step S3 includes the following steps: Step S31: Obtain historical digital trading strategies; Step S32: defining the digital trading strategy state space according to the historical digital trading strategy to obtain the digital trading strategy state space; Step S33: Analyze the transaction strategy correlation characteristics of the historical digital transaction strategies to obtain transaction strategy correlation characteristic data; Step S34: defining the digital trading strategy action space according to the trading strategy associated feature data and the digital trading strategy state space to obtain the digital trading strategy action space; Step S35: Designing a digital trading strategy network according to the digital trading strategy action space to obtain a digital trading strategy network; Step S36: Designing a digital transaction value network according to the digital transaction strategy network and the historical digital transaction strategy to obtain a digital transaction value network; wherein step S36 includes the following steps: Step S361: dividing the historical digital trading strategies into historical trading strategy action nodes according to the digital trading strategy network to obtain historical trading strategy action nodes; Step S362: performing action node status evaluation on the historical transaction strategy action node to obtain historical transaction strategy action node status evaluation data; Step S363: Designing a digital transaction value network according to the historical transaction strategy action node status evaluation data to obtain a digital transaction value network; Step S37: Designing a reinforcement learning model of the digital transaction strategy based on the digital transaction strategy network and the digital transaction value network to establish a digital transaction strategy reinforcement learning model; Step S38: transmitting the weighted digital transaction feature data to the digital transaction strategy reinforcement learning model for digital transaction strategy prediction processing to generate digital transaction strategy prediction data; Step S4: Performing feature derivative analysis and processing of multi-dimensional digital transaction fraud according to the digital transaction strategy prediction data to generate multi-dimensional digital transaction fraud feature derivative data; designing an adversarial game model of abnormal transactions based on the multi-dimensional digital transaction fraud feature derivative data to generate an abnormal transaction adversarial game model; Step S5: performing abnormal transaction adversarial game feature analysis and processing according to the abnormal transaction adversarial game model to generate abnormal transaction adversarial game feature data; establishing a preliminary digital transaction fraud tracing tree model based on the abnormal transaction adversarial game feature data; optimizing the tree structure weight parameters of the preliminary digital transaction fraud tracing tree model according to the digital transaction attention weight information corresponding to the weighted digital transaction feature data to obtain the digital transaction fraud tracing tree model; wherein step S5 includes the following steps: Step S51: performing abnormal transaction adversarial game feature analysis and processing according to the abnormal transaction adversarial game model to generate abnormal transaction adversarial game feature data; wherein step S51 includes the following steps: Step S511: performing abnormal transaction game evolution processing according to the abnormal transaction confrontation game model to generate abnormal transaction game evolution data; Step S512: Performing game evolution replication dynamic processing according to the abnormal transaction game evolution data to generate abnormal transaction game evolution replication dynamic data; Step S513: performing game evolution equilibrium point analysis and processing according to the abnormal transaction game evolution replication dynamic data to generate game evolution equilibrium point data; Step S514: performing abnormal transaction confrontation game feature analysis on the abnormal transaction confrontation game model according to the game evolution equilibrium point data to generate abnormal transaction confrontation game feature data; Step S52: Designing digital transaction fraud tracing tree nodes according to the abnormal transaction adversarial game feature data to obtain digital transaction fraud tracing tree node data; Step S53: Designing the associated labels of the fraud tracing tree structure based on the weighted digital transaction feature data to generate the associated label data of the fraud tracing tree structure; Step S54: setting fraud tracing tree structure weight parameters according to the digital transaction attention weight information corresponding to the weighted digital transaction feature data to obtain fraud tracing tree structure weight parameters; Step S55: designing a preliminary fraud tracing tree model of digital transactions based on the fraud tracing tree structure associated label data for the digital transaction fraud tracing tree node data to establish a preliminary digital transaction fraud tracing tree model; Step S56: optimizing the tree structure weight parameters of the preliminary digital transaction fraud tracing tree model according to the fraud tracing tree structure weight parameters to obtain the digital transaction fraud tracing tree model; Step S6: Analyze and process the weighted digital transaction feature data based on the digital transaction fraud tracing tree model to generate digital transaction abnormality feedback data; transmit the digital transaction abnormality feedback data to the terminal to execute the digital transaction abnormality feedback operation.

2. The method for digital economic transaction processing and anomaly detection based on machine learning according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: collecting digital transaction multimodal data from the digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain digital transaction multimodal data; Step S12: performing digital transaction behavior analysis based on the digital transaction multimodal data to generate digital transaction behavior data; Step S13: performing cluster analysis on the digital transaction behavior data to generate clustered digital transaction behavior data; Step S14: performing transaction pattern recognition processing based on the clustered digital transaction behavior data to generate digital transaction pattern data; Step S15: Design a transaction pattern dimension reduction feature matrix based on the digital transaction pattern data to generate a transaction pattern dimension reduction feature matrix; Step S16: Perform digital transaction feature mapping processing on the digital transaction multimodal data according to the transaction mode dimension reduction feature matrix to generate digital transaction feature data.

3. The method for digital economic transaction processing and anomaly detection based on machine learning according to claim 2 is characterized in that: Step S11 includes the following steps: Collecting initial digital transaction multimodal data on the digital economic transaction platform according to a preset digital economic transaction monitoring engine to obtain initial digital transaction multimodal data; Performing data cleaning on the initial digital transaction multimodal data to obtain cleaned digital transaction multimodal data; Data standardization is performed on the cleaned digital transaction multimodal data to obtain digital transaction multimodal data.

4. The method for digital economic transaction processing and anomaly detection based on machine learning according to claim 2 is characterized in that: Step S2 includes the following steps: Step S21: Perform digital transaction frequency feature analysis based on the digital transaction behavior data to generate digital transaction frequency feature data; Step S22: performing digital transaction time series feature analysis based on the digital transaction multimodal data to generate digital transaction time series feature data; Step S23: formulating digital transaction attention weights according to the digital transaction frequency feature data and the digital transaction time series feature data to obtain digital transaction attention weight information; Step S24: Perform attention weighting processing on the digital transaction feature data according to the digital transaction attention weight information to generate weighted digital transaction feature data.

5. The method for digital economic transaction processing and anomaly detection based on machine learning according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing multi-dimensional digital transaction combination analysis processing according to the digital transaction strategy prediction data to obtain multi-dimensional digital transaction combination data; Step S42: performing a multi-dimensional digital transaction fraud feature analysis based on the multi-dimensional digital transaction combination data to generate multi-dimensional digital transaction fraud feature data; Step S43: using a preset generative adversarial network model to perform fraud feature derivation processing on the multi-dimensional digital transaction fraud feature data to generate multi-dimensional digital transaction fraud feature derived data; Step S44: Design an adversarial game model for abnormal transactions based on the multi-dimensional digital transaction fraud feature derived data to generate an abnormal transaction adversarial game model.

6. The method for digital economic transaction processing and anomaly detection based on machine learning according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: transmitting the weighted digital transaction feature data to the digital transaction fraud tracing tree model for digital transaction anomaly prediction processing to generate digital transaction anomaly prediction data; Step S62: performing digital transaction anomaly assessment processing on the digital transaction anomaly prediction data based on a preset digital transaction anomaly assessment decision to generate digital transaction anomaly assessment data; Step S63: Mapping the digital transaction anomaly assessment data to the digital transaction anomaly prediction data to perform digital transaction anomaly feedback information integration processing to generate digital transaction anomaly feedback data; Step S64: Transmitting the digital transaction abnormality feedback data to the terminal to execute the digital transaction abnormality feedback operation.

Images