Three-party transaction intelligent feedback optimization method and system

Abstract

The invention provides a three-party transaction intelligent feedback optimization method and system. Belongs to the technical field of three-party transaction management, and the method comprises the steps: carrying out the real-time collection of transaction data from a plurality of data sources through a distributed data collection algorithm, and carrying out the preprocessing of the collected transaction data; performing deep feature extraction on the preprocessed data by using a deep learning model, constructing a transaction relation graph by using a graph neural network, and forming a complex transaction network by taking entities as nodes and transaction behaviors as edges; and based on the depth features, performing user sentiment analysis by adopting a depth belief network, and identifying positive, negative or neutral sentiment in user feedback. By integrating various advanced technical means including distributed data acquisition, a deep learning model, a graph neural network, reinforcement learning, a recommendation system, a block chain technology and the like, an intelligent feedback optimization method is provided for three-party transaction.

Description

Technical Field

[0001] The present invention provides a method and system for optimizing intelligent feedback in a three - party transaction, belonging to the technical field of three - party transaction management. Background Art

[0002] Most existing three - party transaction systems rely on simple rules or traditional machine learning models in processing user feedback, making it difficult to capture the deep information and complex relationships in user feedback, resulting in low feedback processing efficiency and prone to ignoring potential fraud behaviors. In addition, traditional methods lack a dynamic optimization mechanism and are difficult to quickly adjust strategies according to real - time transaction data. Summary of the Invention

[0003] The present invention provides a method and system for optimizing intelligent feedback in a three - party transaction to solve the problems mentioned in the above background art:

[0004] A method for optimizing intelligent feedback in a three - party transaction proposed by the present invention, the method includes:

[0005] S1. Real - time collect transaction data from multiple data sources through a distributed data collection algorithm, and pre - process the collected transaction data;

[0006] S2. Apply a deep learning model to extract deep features from the pre - processed data, use a graph neural network to construct a transaction relationship graph, take entities as nodes and transaction behaviors as edges to form a complex transaction network;

[0007] S3. Based on the deep features, use a deep belief network for user sentiment analysis to identify positive, negative or neutral sentiments in user feedback; combine graph neural network and reinforcement learning techniques to construct a dynamic fraud detection model, and detect potential fraud behaviors in real - time by simulating the transaction process;

[0008] S4. Generate personalized product, service or discount recommendations through a deep learning recommendation system combined with user historical behaviors, real - time feedback and the transaction relationship graph; dynamically adjust transaction strategies according to transaction data feedback by applying deep reinforcement learning;

[0009] S5. Based on an intelligent feedback loop mechanism, collect and analyze user feedback in real - time, continuously iterate and optimize recommendations and transaction strategies, and use blockchain technology to record the process of each feedback analysis and strategy adjustment.

[0010] A system for optimizing intelligent feedback in a three - party transaction proposed by the present invention, the system includes:

[0011] A data collection module: Real - time collect transaction data from multiple data sources through a distributed data collection algorithm, and pre - process the collected transaction data;

[0012] Feature Extraction Module: Apply a deep learning model to perform deep feature extraction on the preprocessed data, and use a graph neural network to construct a transaction relationship graph, taking entities as nodes and transaction behaviors as edges to form a complex transaction network;

[0013] Behavior Detection Module: Based on the deep features, use a deep belief network for user sentiment analysis to identify positive, negative, or neutral sentiment in user feedback; Combine graph neural network and reinforcement learning techniques to construct a dynamic fraud detection model, and detect potential fraud behaviors in real time by simulating the transaction process;

[0014] Strategy Adjustment Module: Through a deep learning recommendation system combined with user historical behavior, real-time feedback, and transaction relationship graph, generate personalized product, service, or discount recommendations; According to the feedback of transaction data, apply deep reinforcement learning to dynamically adjust transaction strategies;

[0015] Process Recording Module: Based on an intelligent feedback loop mechanism, collect and analyze user feedback in real time, continuously iterate and optimize recommendations and transaction strategies, and use blockchain technology to record the process of each feedback analysis and strategy adjustment.

[0016] Advantages of the present invention: By integrating a variety of advanced technical means, including distributed data collection, deep learning models, graph neural networks, reinforcement learning, recommendation systems, and blockchain technology, etc., an intelligent feedback optimization method for tripartite transactions is provided. This method can collect and process transaction data in real time, deeply analyze user behavior and market trends, effectively identify fraud behaviors, provide personalized product or service recommendations, and continuously optimize transaction strategies. At the same time, through the application of the intelligent feedback loop mechanism and blockchain technology, the transparency, security, and immutability of data processing are ensured, thereby improving the reliability and user satisfaction of the entire transaction system. Description of the Drawings

[0017] Figure 1 It is a flowchart of the method described in the present invention;

[0018] Figure 2 It is a block diagram of the system described in the present invention. Detailed Embodiments

[0019] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not used to limit the present invention.

[0020] One embodiment of the present invention, as Figure 1 shown, a tripartite transaction intelligent feedback optimization method, the method includes:

[0021] S1. Real-time collect transaction data from multiple data sources (such as e-commerce platforms, financial services, and logistics, etc.) through a distributed data collection algorithm. The transaction data includes transaction records, user behavior logs, product information, and payment details; and preprocess the collected transaction data.

[0022] S2. Apply a deep learning model to extract deep features from the preprocessed data to capture complex features such as user behavior and product attributes; use a graph neural network to construct a transaction relationship graph, with entities as nodes, where the entities include users, products, and transactions, and transaction behaviors as edges to form a complex transaction network.

[0023] S3. Based on the deep features, use a deep belief network for user sentiment analysis to identify positive, negative, or neutral sentiment in user feedback; combine graph neural network and reinforcement learning techniques to construct a dynamic fraud detection model, and detect potential fraud behaviors such as false transactions and money laundering in real time by simulating the transaction process.

[0024] S4. Generate personalized product, service, or offer recommendations through a deep learning recommendation system (such as deep collaborative filtering, neural collaborative filtering) combined with user historical behavior, real-time feedback, and the transaction relationship graph; according to the transaction data feedback, apply deep reinforcement learning to dynamically adjust transaction strategies such as price adjustment and inventory allocation.

[0025] S5. Based on an intelligent feedback loop mechanism, collect and analyze user feedback in real time, continuously iterate and optimize the recommendation and transaction strategies, and use blockchain technology to record the process of each feedback analysis and strategy adjustment.

[0026] The working principle of the above technical solution is as follows: collect data in real time from multiple reliable data sources to ensure the diversity and comprehensiveness of the data. The data sources may include e-commerce platforms, financial service providers, logistics companies, etc.; adopt a distributed data collection algorithm to ensure data synchronization between different data sources and avoid data latency or inconsistency; remove duplicate, invalid or incorrect data, such as null values, outliers, etc. At the same time, perform deduplication and denoising processing on the data to improve data quality; format the data from different sources into a unified format for subsequent processing and analysis. This may include data type conversion, data structure adjustment, etc.; perform standardization or normalization processing on data in different dimensions to ensure their comparability in the model. This helps to reduce biases and errors in the model training process; select an appropriate deep learning model for feature extraction according to the characteristics of the data and business requirements. Commonly used models include convolutional neural networks (CNNs), recurrent neural networks (RNNs), etc.; extract valuable features for subsequent analysis from the original data, such as user behavior features, commodity attribute features, etc. These features can reflect information such as user preferences and market trends; fuse features from different sources to form a more comprehensive and rich feature set. This helps to improve the accuracy and robustness of the model; use users, commodities, and transactions as nodes in the graph. Each node has a unique identifier and attribute information; use transaction behaviors as edges to connect user, commodity, and transaction nodes. The weight of the edge can represent information such as transaction amount and transaction frequency; use graph neural networks to learn the features of nodes and edges and capture potential relationships in the transaction network. This helps to discover social relationships between users, similarities between commodities, etc.; adopt a deep belief network for sentiment analysis of user feedback. By training the model, identify positive, negative, or neutral sentiment in user feedback; generate sentiment labels for each piece of user feedback for subsequent analysis and recommendation. This helps to understand user satisfaction with transactions and commodities and provides a basis for improving services; combine graph neural networks and reinforcement learning techniques to build a dynamic fraud detection model. The model can simulate the transaction process and detect potential fraud behaviors in real time; through training the model, identify fraud behaviors such as false transactions and money laundering. At the same time, the model can continuously optimize the fraud detection strategy according to the detection results to improve the accuracy and efficiency of detection; use a deep learning recommendation system to generate personalized recommendations by combining user historical behavior, real-time feedback, and transaction relationship graph. The recommendation system can consider factors such as user preferences, market trends, and similarities between commodities to provide accurate recommendations; continuously adjust the recommendation strategy according to user feedback and transaction data to optimize the recommendation results. This helps to improve user satisfaction and loyalty; apply deep reinforcement learning to dynamically adjust transaction strategies, such as price adjustment, inventory allocation, etc. The model can continuously optimize the strategy according to transaction data feedback to improve transaction efficiency and profitability; apply the adjusted transaction strategy to actual business and conduct real-time monitoring and evaluation.Adjust the strategy in a timely manner according to the execution effect of the strategy to ensure the steady development of the business; collect user feedback in real time based on the intelligent feedback loop mechanism. This includes information such as users' evaluations and suggestions on transactions, goods, and services; analyze and process the collected feedback to extract valuable information for optimizing recommendation and trading strategies. This helps to continuously improve the business and enhance user satisfaction; use blockchain technology to record the process of each feedback analysis and strategy adjustment. This includes information on all aspects such as feedback collection, processing, analysis, and optimization; the blockchain technology has the characteristic of data immutability, which can ensure the accuracy and reliability of the records. This provides strong support for subsequent auditing and traceability.

[0027] The effects of the above technical solutions are as follows: through the automated and intelligent data collection, processing, and analysis processes, manual intervention and delays are reduced; the application of deep learning models and graph neural networks improves the accuracy of feature extraction and transaction relationship recognition; the personalized recommendation system can generate recommendation content that meets the user's needs based on the user's historical behavior and real-time feedback; the dynamically adjusted trading strategy can better meet the changes in market demand and user needs; the fraud detection model combining deep belief networks and graph neural networks with reinforcement learning can detect potential fraud behaviors in real time; the application of blockchain technology ensures the transparency and traceability of the transaction process and improves the security of transactions; the intelligent feedback loop mechanism can collect and analyze user feedback in real time; continuously iterate and optimize the recommendation and trading strategies to adapt to the changing market environment and user needs.

[0028] In an embodiment of the present invention, the S1 includes:

[0029] S11. Identify and access multiple key data sources, where the data sources include e-commerce platforms, financial service systems, and logistics information systems; use API interfaces to access the data in real time and stably;

[0030] S12. Determine the types of transaction data, where the types include transaction records, user behavior logs (such as browsing, clicking, purchasing), product information (such as name, price, evaluation), and payment details (such as payment method, amount);

[0031] S13. Classify and store the data according to the data type; and preprocess the stored data.

[0032] The working principle of the above technical solution is as follows: First, identify and determine the data sources that are crucial for transaction analysis. These data sources typically cover e-commerce platforms, financial service systems, and logistics information systems. The e-commerce platform provides data such as user purchase behavior and product details; the financial service system contains key data such as payment information and credit assessment; the logistics information system provides data such as order delivery status and logistics timeliness. To ensure the real-time and stability of the data, API (Application Programming Interface) is adopted as the main way of data access. By establishing API connections with each data source, the latest data can be efficiently obtained while ensuring the stability and reliability of data transmission. After the data is accessed, clarify the types of transaction data to be processed according to business requirements. These types include but are not limited to transaction records (recording the specific information of each transaction), user behavior logs (recording various behaviors of users on the platform, such as browsing, clicking, purchasing, etc.), product information (including detailed information such as product name, price, evaluation, etc.), and payment details (involving payment-related information such as payment method and payment amount). Further subdivide each data type for subsequent data processing and analysis. For example, user behavior logs can be subdivided into sub-types such as browsing behavior, clicking behavior, and purchasing behavior. Classify and store the data in different databases or data warehouses according to the data type. This can improve the efficiency of data retrieval and facilitate subsequent data processing and analysis. Perform preprocessing on the stored data, including data cleaning (removing duplicate, invalid, or incorrect data), data formatting (converting the data into a unified format), data standardization / normalization (adjusting the data to the same magnitude for comparison and analysis), etc. The preprocessed data will be cleaner and more standardized, providing a solid foundation for subsequent data analysis and model training.

[0033] The effects of the above technical solutions are as follows: By identifying and accessing multiple key data sources, such as e-commerce platforms, financial service systems, and logistics information systems, the comprehensiveness and diversity of data can be ensured. This helps to obtain a more complete and accurate transaction portrait in subsequent analysis; Using API interfaces for real-time and stable data access can ensure the real-time update of data and the stability of data transmission. This is of great significance for capturing market dynamics and promptly responding to user needs; Defining the types of transaction data, such as transaction records, user behavior logs, commodity information, and payment details, etc., helps with subsequent data processing and analysis work. This can improve the efficiency and accuracy of data processing; Determining the data types according to business requirements can ensure that the collected data is closely related to business goals. This helps to enhance the pertinence and practicality of data analysis; Classifying and storing data according to data types can simplify the data retrieval process and improve the efficiency of data retrieval. This is of great significance for subsequent data analysis and applications; Preprocessing the stored data, such as data cleaning, formatting, and standardization, etc., can improve the quality and usability of data. This helps to reduce data noise and errors and improve the accuracy and reliability of data analysis.

[0034] In one embodiment of the present invention, S2 includes:

[0035] S21. According to business requirements and data characteristics, perform feature selection, and select features that have important impacts on user behavior, commodity attributes, transaction patterns, etc.; Use traditional feature extraction methods (such as statistical methods, text mining, image recognition, etc.) to preliminarily extract the features to form a preliminary feature set;

[0036] S22. According to the preliminarily extracted features and data characteristics, perform feature extraction based on deep learning models (such as convolutional neural network CNN, recurrent neural network RNN, Transformer, graph neural network GNN, etc.), and design the architecture of the deep learning model, including the input layer, hidden layers (such as convolutional layers, recurrent layers, attention layers, graph convolutional layers, etc.), output layer, etc., as well as the connection methods and parameter settings between each layer;

[0037] S23. Train the model, use the trained deep learning model to perform deep feature extraction on the preliminarily extracted features, mine deeper feature information, and fuse the deep features with traditional features to form a comprehensive feature set;

[0038] S24. Perform advanced feature engineering on the fused features, including feature scaling, feature combination, feature selection, etc.; Use feature importance evaluation methods (such as model-based feature importance, statistics-based feature selection, etc.) to screen out key features that have important impacts on the model performance; And perform further processing and optimization on the key features, such as feature dimensionality reduction, feature transformation, etc.

[0039] S25. Construct a transaction relationship graph based on entities such as users, goods, and transactions in the transaction data and the interaction relationships between them; use the entities as nodes and the transaction behaviors as edges to form a complex transaction network;

[0040] S26. Preprocess the transaction relationship graph, such as removing noise nodes and edges, calculating the weights of nodes and edges, etc.; according to the characteristics and requirements of the transaction relationship graph, perform learning and representation through graph neural network models (such as graph convolutional network GCN, graph attention network GAT, graph embedding model, etc.).

[0041] The working principle of the above technical solution is as follows: first, feature selection is performed according to business needs and data characteristics. These features usually include factors that have an important impact on user behavior, product attributes, transaction patterns, etc.; then, the selected features are preliminarily extracted using traditional feature extraction methods (such as statistical methods, text mining, image recognition, etc.). These methods can generate a preliminary feature set based on the statistical properties of the data, text content or image features, etc.; the architecture of the deep learning model is designed based on the preliminarily extracted features and data characteristics. This includes determining the structure of the input layer, hidden layer (such as convolution layer, recurrent layer, attention layer, graph convolution layer, etc.) and output layer, as well as the connection mode and parameter settings between the layers; for example, for image data, a convolutional neural network (CNN) can be used for feature extraction; for sequence data, a recurrent neural network (RNN) or a Transformer model can be used; for graph structure data, a graph neural network (GNN) can be used for feature extraction; the model is trained so that it can learn the deep features of the data. Then, the trained deep learning model is used to perform deep feature extraction on the preliminarily extracted features to mine deeper feature information; the deep features are integrated with traditional features to form a comprehensive feature set. This helps to make full use of the complementarity between different features and improve the performance of the model; perform advanced feature engineering on the fused features, including feature scaling (such as standardization, normalization), feature combination (such as polynomial features, cross features) and feature selection (such as model-based feature importance, statistical-based feature selection, etc.); screen out key features that have a significant impact on model performance, and perform further processing and optimization. This includes feature dimensionality reduction (such as principal component analysis PCA, linear discriminant analysis LDA, etc.) and feature transformation (such as kernel methods, manifold learning, etc.) to simplify model complexity and improve model generalization capabilities; based on entities such as users, products, and transactions in transaction data and the interactions between them, construct a transaction relationship graph; use entities as nodes and transaction behaviors as edges to form a complex transaction network. This helps to reveal the associations and interaction patterns between entities; preprocess the transaction relationship graph, such as removing noise nodes and edges, calculating the weights of nodes and edges, etc. This helps to reduce redundant information and improve the quality of graph data; according to the characteristics and requirements of the transaction relationship graph, select appropriate graph neural network models (such as graph convolutional network GCN, graph attention network GAT, graph embedding model, etc.) for learning and representation. These models can capture the complex relationships and information in the graph structure, providing strong support for subsequent model training and prediction.

[0042] The effects of the above technical solutions are as follows: by selecting features according to business needs and data characteristics, it is possible to ensure that the selected features are closely related to key business aspects such as user behavior, product attributes, and transaction patterns, thereby improving the accuracy and practicality of subsequent analysis and prediction; providing strong support for subsequent deep feature extraction; feature extraction based on deep learning models can dig out deep feature information that is not apparent in the preliminary feature set, thereby improving the richness and expressiveness of features; designing the architecture of the deep learning model according to data characteristics and business needs, including the input layer, hidden layer, and output layer, as well as the connection method and parameter settings between layers, can ensure the efficiency and accuracy of the model when processing different types of data; by training the deep learning model and using it for deep feature extraction, the performance of the model can be significantly improved, including indicators such as accuracy, recall rate, and F1 score; fusing deep features with traditional features can form a comprehensive feature set containing multi-level and multi-dimensional information, providing a richer data foundation for subsequent analysis and prediction; through feature scaling, feature Advanced feature engineering methods such as combination and feature selection can further improve the quality and availability of features and reduce the impact of redundant and noisy information on model performance. The use of feature importance evaluation methods to screen out key features that have an important impact on model performance, and further processing and optimization, such as feature dimension reduction and feature transformation, can simplify the model complexity and improve the model generalization ability. Constructing a transaction relationship graph based on the entities and interaction relationships in the transaction data can present complex transaction behaviors in an intuitive and visual way, which is easy to analyze and understand. Forming a transaction network with entities as nodes and transaction behaviors as edges can reveal the correlation and interaction patterns between entities, providing strong support for subsequent analysis and prediction. Preprocessing the transaction relationship graph, such as removing noise nodes and edges, calculating the weights of nodes and edges, etc., can improve the quality and availability of graph data and provide strong support for subsequent graph neural network learning. Learning and representing the transaction relationship graph through the graph neural network model can capture the complex relationships and information in the graph structure and improve the prediction ability and accuracy of the model.

[0043] In one embodiment of the present invention, the S23 includes:

[0044] Based on the selected deep learning model, a large-scale data set is used for model training. The trained deep learning model is used to perform multi-level and multi-scale deep feature extraction on the initially extracted features.

[0045] Through hidden layers such as convolutional layers, recurrent layers, and attention layers, feature information at different levels is mined, such as local features, global features, and temporal features; and the extracted deep features are preprocessed, including feature normalization and feature smoothing;

[0046] Based on the model-based feature importance evaluation, deep features that have an important impact on subsequent tasks are screened out; the deep features are fused with traditional features through non-linear combination;

[0047] The fused features are further optimized, such as feature dimensionality reduction (e.g., PCA, LDA, etc.), feature transformation (e.g., kernel methods, manifold learning, etc.), and the fused features are screened again through statistics-based feature selection;

[0048] The fused features are verified using a validation dataset to evaluate their performance in subsequent tasks (such as classification, regression, clustering, etc.); according to the verification results, the feature extraction, fusion strategy, and optimization process are iteratively adjusted;

[0049] Features from different modalities (such as text, images, audio, etc.) are fused in a multi-modal manner; a multi-modal deep learning model is used to extract and fuse cross-modal features.

[0050] The working principle of the above technical solution is as follows: According to business requirements and data characteristics, a suitable deep learning model is selected, such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Transformer, or Graph Neural Network (GNN), etc.; The selected deep learning model is trained using a large-scale dataset to ensure that the model can learn the deep features of the data; The trained deep learning model is used to perform multi-level and multi-scale deep feature extraction on the preliminarily extracted features. Through hidden layers such as convolutional layers, recurrent layers, and attention layers, different levels of feature information are mined, such as local features, global features, and temporal features, etc.; The extracted deep features are preprocessed, including feature normalization, feature smoothing, etc., to improve the quality and stability of the features; Based on the feature importance evaluation method of the model, the deep features that have an important impact on subsequent tasks are screened out. For example, a gradient-based feature importance evaluation method can be used to evaluate the importance of features by calculating the gradients of the features on the model output; The deep features and traditional features are fused through non-linear combinations (such as weighted summation, concatenation, etc.) to form a more comprehensive and rich feature set; The fused features are further optimized, such as feature dimensionality reduction (such as Principal Component Analysis PCA, Linear Discriminant Analysis LDA, etc.) and feature transformation (such as kernel methods, manifold learning, etc.), to simplify the model complexity and improve the model generalization ability; Through statistical-based feature selection methods (such as chi-square test, mutual information, etc.), the fused features are screened again to remove redundant and noisy features; The fused features are verified using a validation dataset to evaluate their performance in subsequent tasks (such as classification, regression, clustering, etc.). This can be measured by calculating metrics such as accuracy, recall rate, F1 score, etc.; According to the verification results, the feature extraction, fusion strategy, and optimization process are iteratively adjusted to improve the quality of the features and the performance of the model; For data from different modalities (such as text, image, audio, etc.), a multi-modal deep learning model is used to extract cross-modal features; The extracted multi-modal features are fused to make full use of the complementarity between different modalities. This can be achieved through methods such as concatenation, weighted summation, and attention mechanism.

[0051] The effects of the above technical solutions are as follows: By selecting a deep learning model and training it with a large-scale dataset, multi-level and multi-scale deep features in the preliminary features can be efficiently extracted. These deep features are more expressive and discriminative than traditional features and can more accurately reflect the internal laws and patterns of the data. By using hidden layers such as convolutional layers, recurrent layers, and attention layers, feature information at different levels can be mined, such as local features, global features, and temporal features. This comprehensive way of mining feature information helps subsequent tasks to more accurately understand and utilize the data. Preprocessing the extracted deep features, such as feature normalization and feature smoothing, can improve the quality and stability of the features. At the same time, feature importance evaluation based on the model can screen out deep features that have an important impact on subsequent tasks, further streamline the feature set, and improve the efficiency and performance of the model. By fusing deep features and traditional features through non-linear combination, the advantages of both can be fully utilized to form a more comprehensive and rich feature set. This fusion method helps to improve the generalization ability and adaptability of the model. Further optimizing the fused features, such as feature dimensionality reduction and feature transformation, can simplify the model complexity, improve the model training speed and prediction performance. At the same time, statistic-based feature selection can screen out key features again to further improve the accuracy and robustness of the model. Using the validation dataset to validate the fused features can evaluate their performance in subsequent tasks. According to the validation results, iteratively adjusting the feature extraction, fusion strategy, and optimization process can continuously optimize the feature set and improve the performance of the model. Fusing features from different modalities through multi-modal fusion can make full use of the complementarity between different modalities and improve the comprehensiveness and accuracy of the model. At the same time, using a multi-modal deep learning model to extract and fuse cross-modal features can further mine and utilize the information in multi-modal data and improve the performance and application scope of the model.

[0052] In one embodiment of the present invention, S25 includes:

[0053] Identify key entities such as users, products, and transactions from the transaction data, and classify the entities according to business requirements and data characteristics; Use natural language processing (NLP) techniques, such as named entity recognition (NER), to extract entities from the text data.

[0054] Design attributes for each entity node, such as age, gender, purchasing power, etc. for the user node, category, price, brand, etc. for the product node, and transaction time, transaction amount, etc. for the transaction node, and add timestamps to the node attributes; Identify the interaction behaviors between users and products from the transaction data, such as purchase, browsing, collection, evaluation, etc., and classify them according to the importance and frequency of the behaviors.

[0055] Automatically classify and label transaction behaviors through machine learning algorithms such as behavior pattern recognition, design attributes for each edge, such as the weight of the edge (indicating the importance or frequency of the behavior), the direction of the edge (indicating the initiator and recipient of the behavior), etc., and add additional dimensions to the edge attributes, such as the timestamp of the behavior, the duration of the behavior, the sentiment tendency of the behavior, etc.;

[0056] Connect the identified entity nodes and transaction behavior edges according to business logic and data characteristics to construct a preliminary transaction relationship graph;

[0057] Add global attributes to the transaction relationship graph, such as the density of the graph, the average degree of nodes, the average weight of edges, etc., convert the graph structure into a low-dimensional vector representation through graph embedding techniques such as DeepWalk, Node2Vec, etc.; use graph algorithms such as PageRank, HITS, etc. to evaluate the importance of nodes and edges, and remove noisy nodes and edges;

[0058] According to the attributes of nodes and edges, use a weighted algorithm to calculate the weights of nodes and edges, and dynamically adjust the weights of nodes and edges based on time factors; among them, the weight of the node is calculated by the following formula:

[0059]

[0060] Among them, represents the time decay factor, which means that as time goes by, the weight of the behavior will gradually decrease;

[0061] The weight of the edge is calculated by the following formula:

[0062]

[0063] Among them, f i represents the frequency of behavior i; w i represents the fixed weight of behavior i (predetermined according to the importance of the behavior); δt i represents the difference between the current time and the time t when the behavior occurred i ; α represents the time decay coefficient; s i represents the sentiment tendency score of behavior i (for example, a positive evaluation may be a positive value, and a negative evaluation is a negative value);

[0064] Introduce the time dimension, construct a dynamic transaction relationship graph, use time series analysis techniques such as ARIMA, LSTM, etc. to perform time series prediction and trend analysis on the transaction relationship graph, and adjust and optimize the transaction relationship graph based on the analysis results.

[0065] The working principle of the above technical solution is as follows: Identify key entities such as users, commodities, and transactions from transaction data. This is the basis for constructing a transaction relationship graph; Classify the identified entities according to business requirements and data characteristics. For example, users can be classified according to purchasing power, age, gender, etc.; Commodities can be classified according to category, price, brand, etc.; Use natural language processing (NLP) techniques, such as named entity recognition (NER), to extract entities from text data. This helps to extract useful information from unstructured data and enrich the content of the transaction relationship graph; Design attributes for each entity node, such as the age, gender, purchasing power, etc. of the user node; The category, price, brand, etc. of the commodity node; The transaction time, transaction amount, etc. of the transaction node. These attributes help to more comprehensively describe the characteristics of the entity nodes; Identify the interaction behaviors between users and commodities from transaction data, such as purchase, browse, favorite, evaluation, etc.; Classify the identified interaction behaviors according to the importance and frequency of the behaviors. This helps to distinguish the impact degree of different behaviors on the transaction relationship graph; Use machine learning algorithms, such as behavior pattern recognition, to automatically classify and label transaction behaviors. Design attributes for each edge, such as the weight of the edge (indicating the importance or frequency of the behavior), the direction of the edge (indicating the initiator and receiver of the behavior), etc. At the same time, add additional dimensions to the edge attributes, such as the timestamp of the behavior, the duration of the behavior, the sentiment tendency of the behavior, etc. These attributes help to more accurately describe transaction behaviors; Connect the identified entity nodes and transaction behavior edges according to business logic and data characteristics to construct a preliminary transaction relationship graph. This is the basis for subsequent analysis and prediction; Add global attributes to the transaction relationship graph, such as the density of the graph, the average degree of nodes, the average weight of edges, etc. These attributes help to understand the overall characteristics of the transaction relationship graph; Use graph embedding techniques, such as DeepWalk, Node2Vec, etc., to convert the graph structure into a low-dimensional vector representation. This helps to reduce the computational complexity and improve the analysis efficiency; Use graph algorithms, such as PageRank, HITS, etc., to evaluate the importance of nodes and edges. According to the evaluation results, remove noise nodes and edges to improve the quality of the transaction relationship graph; Calculate the weights of nodes and edges using a weighted algorithm according to the attributes of nodes and edges. At the same time, dynamically adjust the weights of nodes and edges based on time factors. This helps to reflect the changing trends of entities and behaviors in the transaction relationship graph; Introduce the time dimension to construct a dynamic transaction relationship graph. This helps to analyze the evolution process of the transaction relationship graph over time; Use time series analysis techniques, such as ARIMA, LSTM, etc., to perform time series prediction and trend analysis on the transaction relationship graph. Based on the analysis results, adjust and optimize the transaction relationship graph. This helps to predict future transaction trends and provide a basis for business decisions.

[0066] The effects of the above technical solutions are as follows: By accurately identifying key entities such as users, goods, and transactions from transaction data and classifying them according to business requirements and data characteristics, a solid foundation is provided for subsequent analysis and modeling. This helps to ensure the accuracy and consistency of data and improve the reliability of subsequent analysis; Using natural language processing (NLP) techniques, such as named entity recognition (NER), to extract entities from text data can efficiently extract useful information. This helps to make full use of valuable information in unstructured data and enrich the content of the transaction relationship graph; Design comprehensive attributes for each entity node, such as age, gender, purchasing power, etc. for the user node, category, price, brand, etc. for the goods node, transaction time, transaction amount, etc. for the transaction node, and add timestamps. This helps to more comprehensively describe the characteristics of entity nodes and improve the accuracy and robustness of the model; Identify the interaction behaviors between users and goods from transaction data and classify them according to the importance and frequency of the behaviors. This helps to more precisely understand users' behavior habits and preferences and provide strong support for subsequent analysis and prediction; Automatically classify and label transaction behaviors through machine learning algorithms, design attributes for each edge, and add additional dimensions. This helps to improve the automation level of analysis, reduce manual intervention, and improve efficiency and accuracy; Connect the identified entity nodes and transaction behavior edges according to business logic and data characteristics to construct a preliminary transaction relationship graph. This helps to intuitively display the relationships between users, goods, and transactions and provide a basis for subsequent analysis and decision-making; Add global attributes to the transaction relationship graph and convert the graph structure into a low-dimensional vector representation through graph embedding technology. This helps to reduce the dimensionality of data, improve computational efficiency, and retain important information in the graph structure; Use graph algorithms to evaluate the importance of nodes and edges and remove noisy nodes and edges. This helps to purify data and improve the accuracy and robustness of the model; According to the attributes of nodes and edges, use a weighted algorithm to calculate weights and dynamically adjust them based on time factors. This helps to reflect the changing trend of data over time and improve the timeliness and accuracy of the model; Introduce the time dimension, construct a dynamic transaction relationship graph, and use time series analysis techniques for prediction and trend analysis. This helps to predict future transaction trends and provide strong support for business decisions. The above formula includes the fixed weight of the behavior, behavior frequency, time decay factor, and sentiment tendency score. These factors work together in weight calculation, enabling the final weight value to more comprehensively reflect the actual importance of nodes and edges. By introducing the time decay factor, the formula can automatically adjust the weight according to the time when the behavior occurs. As time goes by, the weight of early behaviors will gradually decrease, while the weight of recent behaviors will be relatively higher. This dynamic adjustment mechanism helps to ensure the timeliness and accuracy of network analysis. The introduction of behavior frequency and sentiment tendency score enables weight calculation to reflect users' actual usage of nodes and edges and their sentiment tendency.This helps to understand user behavior and network structure more deeply. The fixed weights and time decay coefficients in the formula can be adjusted according to the actual situation. This enables the algorithm to flexibly adapt to different application scenarios and data characteristics, thereby improving the practicality and adaptability of the algorithm; this weighted algorithm is applicable to networks with complex structures and dynamic characteristics. By calculating the weights of nodes and edges, key information such as key nodes, important paths, and potential influence propagation paths in the network can be revealed. This is of great significance for fields such as social network analysis, recommendation systems, and information dissemination.

[0067] In one embodiment of the present invention, S3 includes:

[0068] S31. Perform sentiment analysis on user feedback through a deep belief network (DBN) to identify positive, negative, or neutral sentiment; and use the sentiment analysis results to evaluate user satisfaction;

[0069] S32. Combine graph neural network (GNN) and reinforcement learning (RL) technologies to build a dynamic fraud detection model; and set fraud detection strategies, such as simulating the trading process and detecting abnormal trading behaviors (such as frequent large - amount transactions, false evaluations, etc.);

[0070] S33. Use reinforcement learning algorithms (such as Q - learning, DeepQ - Network, etc.) to train the dynamic fraud detection model and perform real - time detection of potential fraud behaviors, such as false transactions, money laundering, etc.;

[0071] S34. Mark and record the identified fraud behaviors, and take corresponding response measures (such as warnings, freezing accounts, reporting to the police, etc.) according to the nature and severity of the fraud behaviors.

[0072] The working principle of the above technical solution is as follows: The deep belief network is a multi-layer neural network that can learn the deep features of data. In this step, the DBN is used for sentiment analysis of user feedback to identify positive, negative, or neutral sentiment; the result of the sentiment analysis is used to evaluate user satisfaction. By analyzing the sentiment tendency in user feedback, the satisfaction of users with products or services can be indirectly understood, providing a basis for improving products or services; The graph neural network is good at processing graph-structured data and can capture the relationships between nodes and edges. In this step, the GNN is used to construct a transaction relationship graph to analyze the associations between users, goods, and transactions; Reinforcement learning is a method of learning optimal strategies through trial and error. In this step, RL is used to construct a dynamic fraud detection model to optimize the fraud detection strategy by simulating the transaction process and learning abnormal transaction behaviors; The fraud detection strategy includes simulating the transaction process and detecting abnormal transaction behaviors (such as frequent large-value transactions, false evaluations, etc.). These strategies aim to identify potential fraud behaviors and ensure the security and fairness of transactions; Reinforcement learning algorithms (such as Q-learning, DeepQ-Network, etc.) are used to train the dynamic fraud detection model. Through continuous trial and error and learning, the model can gradually optimize the fraud detection strategy and improve the accuracy and efficiency of detection; The trained model can perform real-time detection of potential fraud behaviors, such as false transactions, money laundering, etc. This helps to promptly discover and handle fraud behaviors, protecting the property safety of users and the reputation of the trading platform; The identified fraud behaviors are marked and recorded for subsequent analysis and processing. This helps to establish a fraud behavior database, providing experience and reference for future fraud detection; According to the nature and severity of the fraud behavior, corresponding response measures are taken. These measures may include warnings, freezing accounts, reporting to the police, etc., aiming to promptly stop fraud behaviors and protect the legitimate rights and interests of users.

[0073] The effects of the above technical solutions are as follows: By performing sentiment analysis on user feedback through a deep belief network (DBN), it is possible to accurately identify positive, negative, or neutral sentiments, thereby providing users with a more personalized service experience. At the same time, the sentiment analysis results can be used to evaluate user satisfaction, helping enterprises promptly understand users' true feelings towards products or services, and providing a strong basis for product improvement and service optimization; By combining graph neural network (GNN) and reinforcement learning (RL) technologies, the constructed dynamic fraud detection model can capture abnormal behaviors during the transaction process in real time, such as frequent large-value transactions, false evaluations, etc. This model not only improves the accuracy and efficiency of fraud detection but also can adapt to continuously changing fraud means, ensuring the security and reliability of the transaction environment; Using reinforcement learning algorithms (such as Q-learning, DeepQ-Network, etc.) to train the dynamic fraud detection model enables it to have the ability of self-learning and optimization. This can not only achieve intelligent identification of potential fraud behaviors but also perform real-time detection, effectively curbing the occurrence of fraud behaviors and protecting users' property safety; Marking and recording the identified fraud behaviors helps to establish a fraud behavior database, providing data support for subsequent analysis and prediction. At the same time, according to the nature and severity of fraud behaviors, taking corresponding response measures, such as warnings, freezing accounts, reporting to the police, etc., can promptly and effectively handle fraud incidents, maintaining the order of the trading platform and the legitimate rights and interests of users; Through accurate sentiment analysis and satisfaction evaluation, as well as efficient fraud detection and response measures, it is possible to significantly improve users' satisfaction and trust in the trading platform. This helps to enhance user loyalty and stickiness, promoting the long-term development of the trading platform; Effective fraud detection and prevention measures can significantly reduce the economic losses and operating costs incurred by enterprises due to fraud behaviors. At the same time, by improving user satisfaction and trust, it is possible to reduce user churn and negative word-of-mouth, reducing the market risk and brand reputation risk of enterprises.

[0074] In one embodiment of the present invention, step S4 includes:

[0075] S41. Fusing user historical behavior data (such as browsing, clicking, purchase records), real-time feedback data (such as evaluations, ratings), and transaction relationship graph data (such as user-product interactions, product-product similarities) through a fusion algorithm to form a comprehensive user profile;

[0076] S42. Performing feature engineering on the fused data and extracting features; extracting features that have important impacts on user preferences, product attributes, transaction patterns, etc.; constructing a recommendation system through a deep learning model according to business requirements and data characteristics; and designing various recommendation strategies, such as popular product recommendations, similar product recommendations, personalized discount recommendations, and recommendations based on users' social networks;

[0077] S43. Evaluate and optimize the recommendation strategy through A / B testing, and select the optimal recommendation strategy; conduct in-depth analysis of transaction data to extract indicators that have important impacts on transaction strategy adjustment, such as user purchase behavior, product sales volume, inventory situation, and market trends. Among them, the effect of the recommendation strategy is evaluated by the following formula:

[0078]

[0079] Among them, E(S) represents the effect evaluation score of the recommendation strategy S; N represents the number of positive impact indicators; f i (S) represents the calculation function of the i-th positive impact indicator, such as click-through rate (CTR), conversion rate (CVR), average order value (AOV), etc.; w i represents the weight of the i-th positive impact indicator; M represents the number of adjustment factors for positive impact indicators; g j (S) represents the j-th adjustment factor, which can be time decay, user activity, market trend, etc.; β ij represents the interaction coefficient between the i-th indicator and the j-th adjustment factor; L represents the number of negative impact indicators; h k (S) represents the calculation function of the k-th negative impact indicator, such as user complaint rate, recommendation deviation, system error rate, etc.; u k represents the weight of the k-th negative impact indicator; P represents the number of adjustment factors for negative impact indicators; p l (S) represents the l-th negative impact adjustment factor; γ kl represents the interaction coefficient between the k-th negative impact indicator and the l-th adjustment factor;

[0080] S44. Use data mining techniques (such as association rule mining, clustering analysis, time series analysis, etc.) to mine potential laws and patterns in transaction data, and construct a deep reinforcement learning (DRL) model based on the results of transaction data analysis for dynamically adjusting transaction strategies; use historical transaction data for model training and continuously iterate and optimize;

[0081] S45. Apply the decision results of the DRL model to the transaction strategy adjustment algorithm for intelligent and automated adjustment of transaction strategies; the measurement adjustment algorithm includes a price adjustment algorithm (adjusting prices based on factors such as market demand, competition situation, and cost changes), an inventory allocation algorithm (allocating inventory based on factors such as product sales volume, inventory level, and transportation cost), and a promotion strategy algorithm (formulating promotion strategies based on factors such as user preferences, market trends, and holidays);

[0082] S46. Evaluate the effectiveness of the adjusted trading strategy. By comparing indicators such as sales volume, user satisfaction, and inventory turnover before and after the adjustment, evaluate the effect of the strategy adjustment; feedback the evaluation results to the DRL model to guide the further optimization and training of the model.

[0083] The working principle of the above technical solution is as follows: First, the user's historical behavior data (such as browsing, clicking, and purchase records), real-time feedback data (such as evaluations and ratings), and transaction relationship graph data (such as user-product interaction and product-product similarity) are integrated through a fusion algorithm. This step aims to integrate data from different sources to form a comprehensive understanding of user behavior; the integrated data is used to build a comprehensive user portrait. The user portrait contains information such as user preferences, interests, and behavior patterns, providing a basis for the construction of the subsequent recommendation system; feature engineering is performed on the integrated data to extract features that have a significant impact on user preferences, product attributes, transaction patterns, etc.; based on business needs and data characteristics, a deep learning model is used to build a recommendation system. This step aims to use the powerful learning ability of the deep learning model to mine the potential relationship between users and products from the data; design a variety of recommendation strategies, such as hot product recommendations, similar product recommendations, personalized discount recommendations, and recommendations based on user social networks. These strategies are designed to meet the needs and preferences of different users and improve the accuracy and satisfaction of recommendations; the recommendation strategies are evaluated and optimized through A / B testing. A / B testing is a commonly used experimental design method used to compare the effects of different strategies and select the optimal strategy; conduct in-depth analysis of transaction data to extract indicators that have an important impact on the adjustment of transaction strategies. These indicators include user purchasing behavior, product sales, inventory status, market trends, etc., providing data support for subsequent transaction strategy adjustments; use data mining techniques (such as association rule mining, cluster analysis, time series analysis, etc.) to mine potential laws and patterns in transaction data. These laws and patterns help understand market trends and user behavior, and provide a basis for subsequent transaction strategy adjustments; build a deep reinforcement learning (DRL) model based on the results of transaction data analysis. The DRL model can learn the mapping relationship between transaction strategies and results, and train and optimize the model based on historical transaction data; apply the decision results of the DRL model to the transaction strategy adjustment algorithm. These algorithms include price adjustment algorithms, inventory allocation algorithms, promotion strategy algorithms, etc., which aim to achieve intelligent and automated adjustments to transaction strategies; according to the decision results of the DRL model, intelligent and automated adjustments are made to transaction strategies. These adjustments are aimed at optimizing transaction results and improving indicators such as sales, user satisfaction, and inventory turnover; and evaluate the effectiveness of the adjusted transaction strategies. By comparing sales, customer satisfaction, inventory turnover and other indicators before and after the adjustment, the effect of the strategy adjustment is evaluated; the evaluation results are fed back to the DRL model to guide the further optimization and training of the model. This step aims to improve the decision-making ability of the DRL model and the adjustment effect of the trading strategy through continuous iteration and optimization.

[0084] The effects of the above technical solutions are as follows: By integrating the user's historical behavior data, real-time feedback data, and transaction relationship graph data through a fusion algorithm, a comprehensive and accurate user profile is formed. This helps to deeply understand user preferences, needs, and behavior patterns, providing strong support for the formulation of subsequent recommendation and transaction strategies; Using a deep learning model to construct a recommendation system and combining multiple recommendation strategies (such as popular product recommendations, similar product recommendations, personalized discount recommendations, and recommendations based on the user's social network) can provide users with more personalized and accurate recommendation services. This not only improves the user experience but also promotes product sales and platform activity; Through A / B testing, the effectiveness of the recommendation strategy is evaluated to ensure the selection of the optimal strategy. At the same time, in-depth analysis of transaction data is carried out to extract key indicators, providing a scientific basis for the adjustment of transaction strategies. This scientific and rigorous method helps to avoid blind decision-making and improve the effectiveness and accuracy of strategy adjustment; Using data mining techniques to mine potential laws and patterns in transaction data and constructing a deep reinforcement learning (DRL) model to achieve dynamic adjustment of transaction strategies. The DRL model can adjust strategies in real time according to market changes and user behavior, improving transaction efficiency and effectiveness. This dynamic and flexible adjustment method helps enterprises quickly adapt to market changes and maintain a competitive advantage; Applying the decision results of the DRL model to the transaction strategy adjustment algorithm to achieve intelligent and automated adjustment of transaction strategies. This reduces the cost and error of manual intervention and improves the efficiency and accuracy of transaction management. At the same time, intelligent and automated transaction management helps enterprises achieve large-scale and efficient operations; Evaluating the effectiveness of the adjusted transaction strategy and feeding back the evaluation results to the DRL model to guide the further optimization and training of the model. This ability to continuously optimize and iterate helps enterprises continuously improve the effectiveness and accuracy of transaction strategies and maintain market competitiveness; Through accurate recommendations and dynamic adjustment of transaction strategies, the personalized needs of users can be met, improving user satisfaction and loyalty. This helps enterprises establish a stable user group and promote long-term development; Intelligent and automated transaction management and precise strategy adjustment help to reduce operating costs, reduce inventory backlogs and waste. At the same time, by mining potential laws and patterns in transaction data, enterprises can timely discover market trends and potential risks, providing strong support for decision-making and reducing operating risks. The above formula includes positive impact indicators (such as click-through rate, conversion rate, average order value, etc.) and negative impact indicators (such as user complaint rate, recommendation deviation, system error rate, etc.), which can comprehensively reflect the performance of the recommendation strategy in multiple dimensions. By introducing positive and negative impact adjustment factors (such as time decay, user activity, market trend, etc.), the formula can more carefully consider the impact of different factors on the recommendation effect, making the evaluation result more accurate. The weights in the formula allow for flexible adjustment according to business needs and actual situations to highlight the importance of certain key indicators.The interaction coefficient can reflect the complex relationships between different indicators and regulatory factors, making the evaluation model more in line with the actual business scenario; by continuously iterating and optimizing the parameters in the formula, the effectiveness of the recommendation strategy can be continuously improved to meet the changing market demands and user preferences. The formula is based on transaction data and user behavior data, and obtains the effectiveness evaluation score of the recommendation strategy through quantitative analysis, providing strong support for data-driven decision-making. By comparing the effectiveness evaluation scores of different recommendation strategies, the optimal strategy can be objectively selected, avoiding subjective assumptions and blind decisions. By improving positive impact indicators such as click-through rate, conversion rate, and average order value, as well as reducing negative impact indicators such as user complaint rate and recommendation bias, the business benefits and profitability can be significantly improved.

[0085] In one embodiment of the present invention, the S44 includes:

[0086] Through association rule mining algorithms such as Apriori and FP-Growth, the association relationships between products are mined from transaction data, such as "users who buy milk often also buy bread", and the mined association rules are visually displayed to help business personnel understand the relevance between products and the user purchase behavior patterns;

[0087] Using clustering algorithms such as K-means and DBSCAN, users or products are clustered to discover potential user groups or product categories, and the clustering results are interpreted and analyzed to extract the characteristics of the cluster centers;

[0088] Through time series analysis algorithms, time series prediction is performed on transaction data to predict future sales trends and changes in user behavior; combined with relevant factors, such as seasonal factors, holiday effects, etc., the time series prediction results are corrected and optimized;

[0089] According to the requirements of transaction strategy adjustment and data characteristics, the architecture of a deep reinforcement learning model is constructed, including an input layer, hidden layers (such as convolutional layers, recurrent layers, attention layers, etc.), an output layer, etc.; the state space is defined, including key information such as user behavior, product attributes, and market trends in transaction data; and the action space is designed, including transaction strategy adjustment actions such as price adjustment, inventory allocation, and promotion strategies.

[0090] According to the business objectives, a reward function is designed, considering multiple dimensions such as sales volume, user satisfaction, and inventory turnover rate, weighted and combined for the reward function, and historical transaction data is used for model training, and the model parameters are updated through iterative optimization algorithms (such as gradient descent, Adam, etc.);

[0091] Through a simulation system, the model is trained and tested, and according to the training results and test feedback, the model architecture, state space, action space, and reward function are adjusted and optimized.

[0092] The working principle of the above technical solution is as follows: Use association rule mining algorithms such as Apriori and FP-Growth to traverse the transaction data set to find frequent itemsets (i.e., combinations of goods that often appear together); Based on the frequent itemsets, generate association rules and calculate their confidence levels (i.e., conditional probabilities, representing the probability of the subsequent item appearing given the previous item); Filter out the association rules with confidence levels higher than the preset threshold, and these rules reveal the associations between goods; Display the mined association rules in a graphical manner, such as a commodity association network diagram, to help business personnel intuitively understand the associations between goods and the user purchase behavior patterns; Use clustering algorithms such as K-means and DBSCAN to cluster user or commodity feature data; The algorithm divides users or commodities into different groups or categories according to the similarity of feature data (such as user purchase behavior, commodity attributes, etc.); Extract the features of the cluster centers, representing the common features or trends of the group; Interpret and analyze the clustering results to reveal potential user groups or commodity categories, providing a basis for precision marketing and commodity recommendation; Use time series analysis algorithms (such as ARIMA, LSTM, etc.) to perform time series prediction on transaction data; Predict future sales trends and changes in user behavior, providing forward-looking guidance for transaction strategy adjustment; Combine relevant factors such as seasonal factors and holiday effects to correct and optimize the time series prediction results, improving the accuracy and reliability of the prediction; According to the requirements of transaction strategy adjustment and data characteristics, construct the architecture of a deep reinforcement learning model; The model includes an input layer (receiving transaction data), hidden layers (such as convolutional layers, recurrent layers, attention layers, etc., for feature extraction and pattern recognition), and an output layer (outputting transaction strategy adjustment actions); Define the state space, including key information such as user behavior, commodity attributes, and market trends in transaction data; Design the action space, including transaction strategy adjustment actions such as price adjustment, inventory allocation, and promotion strategies; According to the business objectives, design a reward function, considering multiple dimensions such as sales volume, user satisfaction, and inventory turnover rate; Perform weighted sum and combination on the reward function to reflect the priorities and trade-offs between different objectives; Use historical transaction data for model training, and update the model parameters through iterative optimization algorithms (such as gradient descent, Adam, etc.); Construct a simulation system to train and test the model; According to the training results and test feedback, adjust and optimize the model architecture, state space, action space, and reward function; Through continuous iteration and optimization, improve the decision-making ability of the model and the effect of transaction strategy adjustment.

[0093] The effects of the above technical solutions are as follows: By using algorithms such as Apriori and FP-Growth, the association relationships between products are mined, such as "users who buy milk often also buy bread". This kind of insight helps enterprises understand the user purchase behavior patterns, optimize product combinations and recommendation strategies, and improve sales and user satisfaction; Visualize the mined association rules, enabling business personnel to intuitively understand the associations between products and user behavior patterns, facilitating quick decision-making and adjustment; Use clustering algorithms such as K-means and DBSCAN to cluster users or products, discovering potential user groups or product categories. This helps enterprises more accurately target target users and market segments, and formulate personalized marketing strategies and product recommendations; Interpret and analyze the clustering results, extract the characteristics of the cluster centers, and provide enterprises with deeper market insights and user portraits; Through time series analysis algorithms, predict transaction data to reveal future sales trends and changes in user behavior. This helps enterprises plan ahead, optimize inventory management and production plans, and reduce the risks of inventory backlogs and out-of-stock; Combine relevant factors such as seasonal factors and holiday effects to correct and optimize the time series prediction results, improving the accuracy and reliability of the prediction; According to the needs of transaction strategy adjustment and data characteristics, construct a deep reinforcement learning model to achieve intelligent and automated adjustment of transaction strategies. The model can dynamically adjust prices, inventory, and promotion strategies according to real-time market feedback and changes in user behavior, improving transaction efficiency and effectiveness; According to business goals, design a reward function, comprehensively considering multiple dimensions such as sales, user satisfaction, and inventory turnover rate. This helps enterprises balance user experience and operational efficiency while pursuing economic benefits; Use historical transaction data for model training, and update model parameters through iterative optimization algorithms. At the same time, train and test the model through a simulation system, and make adjustments and optimizations according to the training results and test feedback to ensure the accuracy and stability of the model; The comprehensive application of the above technical solutions can significantly enhance the market competitiveness and market response speed of enterprises. By deeply mining data value and optimizing transaction strategies, enterprises can better meet user needs, improve user satisfaction and loyalty, and thus achieve growth in sales and market share; The technical solutions have the ability to continuously iterate and optimize, and can be continuously adjusted and upgraded with changes in the market environment and user needs, ensuring that enterprises always maintain a competitive advantage.

[0094] In one embodiment of the present invention, S5 includes:

[0095] S51. Integrate feedback data from multiple channels such as user behavior, transaction records, system logs, and third-party data sources, and preprocess the collected feedback data. The preprocessing includes cleaning, removing noise, outliers, and duplicate data, and at the same time performing unified and standardized processing of data formats;

[0096] S52. Extract key features from the feedback data based on machine learning or deep learning techniques, and perform encoding and transformation. Through clustering and association rule mining, analyze user behavior patterns and transaction characteristics, and identify potential market trends and user needs;

[0097] S53. Use natural language processing techniques to perform sentiment analysis on text data such as user evaluations and comments, and understand users' satisfaction and opinions on products and services;

[0098] S54. Based on historical data and features, build a prediction model to predict users' future behavior trends and transaction intentions. Through a real-time feedback system, instantly capture and process user feedback; through iterative optimization of the feedback loop, continuously adjust and optimize the processes of collecting, processing, analyzing, and applying feedback data;

[0099] S55. According to the feedback analysis results, intelligently adjust the recommendation algorithm and trading strategy. According to business requirements and data characteristics, select a blockchain consensus mechanism (such as proof of work, proof of stake, practical Byzantine fault tolerance, etc.); and deploy a blockchain network;

[0100] S56. Upload key information such as feedback data and transaction records to the blockchain through a smart contract. Utilize the immutability of the blockchain to verify and audit the data, and based on the data access control mechanism, enable only authorized users or systems to access and modify the data on the blockchain.

[0101] The working principle of the above technical solution is as follows: collect feedback data from multiple channels such as user behavior, transaction records, system logs, and third-party data sources; clean the collected data to remove noise (such as invalid information), outliers (such as extreme or unreasonable data), and duplicate data to ensure data quality; perform unified and standardized processing on the data format to make data from different sources compatible and integrable; extract key features from the preprocessed feedback data based on machine learning or deep learning techniques; encode and transform the features for subsequent analysis; divide user or transaction data into different groups or categories through clustering analysis to reveal user behavior patterns and transaction characteristics; apply association rule mining techniques to discover the correlation between products or services and identify potential market trends and user needs; use natural language processing techniques to perform sentiment analysis on text data such as user evaluations and comments; judge the satisfaction and opinion tendency (positive, negative, or neutral) of users towards products and services through methods such as sentiment dictionaries and machine learning models; the sentiment analysis results can be used to improve products or services and enhance user satisfaction; build prediction models (such as time series models, classification models, regression models, etc.) based on historical data and extracted features; predict users' future behavior trends and transaction intentions to provide a basis for adjusting recommendation algorithms and transaction strategies; capture and process user feedback in real time through a real-time feedback system to ensure the timeliness and accuracy of data; continuously adjust and optimize the processes of collecting, processing, analyzing, and applying feedback data through an iterative optimization feedback loop to improve the accuracy and generalization ability of the prediction model; intelligently adjust recommendation algorithms and transaction strategies according to the feedback analysis results to better meet user needs and market changes; select an appropriate blockchain consensus mechanism (such as proof of work, proof of stake, practical Byzantine fault tolerance, etc.) according to business requirements and data characteristics and deploy a blockchain network; achieve decentralized storage and verification of data through blockchain technology to ensure data security and credibility; upload key information such as feedback data and transaction records to the blockchain through smart contracts and use the immutability of the blockchain to verify and audit the data; ensure that only authorized users or systems can access and modify the data on the blockchain based on a data access control mechanism (such as permission control in smart contracts, encryption technology, etc.); achieve data transparency and traceability through blockchain technology for easy data management and auditing.

[0102] The effects of the above technical solution are as follows: By integrating feedback data from multiple channels, cleaning and preprocessing it, noise, outliers, and duplicate data are removed, ensuring the accuracy and reliability of the data; unifying the data format and standardizing the processing enables data from different sources to be compatible and integrated, facilitating subsequent analysis; based on machine learning or deep learning techniques, key features are extracted from the feedback data, and through clustering and association rule mining, user behavior patterns and transaction characteristics are revealed, helping to identify potential market trends and user needs; using natural language processing techniques, sentiment analysis is performed on text data such as user evaluations and comments to understand users' satisfaction and opinions on products and services, providing important references for product improvement and customer service; based on historical data and features, a prediction model is constructed, which can predict users' future behavior trends and transaction intentions, providing a scientific basis for adjusting recommendation algorithms and transaction strategies; through a real-time feedback system, user feedback is captured and processed immediately, and through an iterative optimization feedback loop, the processes of collecting, processing, analyzing, and applying feedback data are continuously adjusted and optimized, improving the accuracy of prediction and the timeliness of decision-making; according to the feedback analysis results, the recommendation algorithm and transaction strategy are intelligently adjusted to better meet user needs and market changes, enhancing market competitiveness; according to business requirements and data characteristics, a suitable blockchain consensus mechanism is selected, and a blockchain network is deployed to achieve decentralized storage and verification of data, improving data security and credibility; key information such as feedback data and transaction records is uploaded to the blockchain through smart contracts, and using the immutability of the blockchain, the data is verified and audited to ensure the authenticity and integrity of the data; based on the data access control mechanism, only authorized users or systems can access and modify the data on the blockchain, ensuring data security and privacy; by deeply mining user needs and feedback, products and services are optimized, improving user satisfaction and loyalty; by intelligently adjusting strategies and uploading data to the blockchain, the operation process is optimized, improving operation efficiency and market response speed; through data verification and audit, the risk of data leakage and tampering is reduced, providing a more secure and reliable operation environment for the enterprise.

[0103] In one embodiment of the present invention, the S52 includes:

[0104] Extract multi-dimensional features from user behavior data (such as browsing, clicking, purchase records), transaction records, system logs, and third-party data sources, including user attributes (age, gender, region), behavior characteristics (access frequency, stay time, purchase conversion rate), transaction characteristics (order amount, purchase category, purchase frequency), system performance (response time, error rate), etc.; perform importance evaluation on the extracted features, and use feature selection algorithms (such as recursive feature elimination, Lasso regression) to screen out features that have a significant impact on the analysis target;

[0105] Perform one-hot encoding or label encoding on categorical features, and perform standardization (such as Z-score standardization) or normalization on numerical features;

[0106] Capture the non-linear relationships between features through polynomial feature generation or feature crossing methods; Use deep learning techniques (such as embedding layers) for dimensionality reduction and representation learning of high-dimensional sparse features;

[0107] Use the K-means algorithm to perform clustering analysis on user behavior data, identify user groups with similar behavior patterns, and visually display the clustering results. For example, use t-SNE or PCA for dimensionality reduction and then draw two-dimensional or three-dimensional scatter plots to help business personnel understand the distribution and differences of user behavior;

[0108] Perform clustering analysis on transaction records to identify user groups or product categories with similar transaction characteristics; Combine the clustering results and business knowledge to interpret and analyze the cluster centers and extract valuable insights;

[0109] Through association rule mining algorithms, mine the association relationships between products from transaction data, such as "users who buy milk often also buy bread"; Combine user behavior data and transaction records to mine the association relationships between user behaviors, such as "users who have browsed a certain type of product are more likely to buy another type of product";

[0110] Capture the patterns and trends in user behavior sequences by constructing user behavior sequence models (such as hidden Markov models, recurrent neural networks).

[0111] The working principle of the above technical solution is as follows: Extract multi-dimensional features from user behavior data (such as browsing, clicking, purchase records, etc.), transaction records, system logs, and third-party data sources; Use feature selection algorithms (such as recursive feature elimination, Lasso regression) to evaluate the importance of the extracted features and screen out the features that have a significant impact on the analysis target. This step helps to reduce noise and improve the accuracy and efficiency of subsequent analysis; Perform one-hot encoding (One-Hot Encoding) or label encoding (Label Encoding) on categorical features so that machine learning models can handle non-numerical data; Normalize (such as Z-score normalization) or standardize numerical features to ensure that different features are on the same numerical scale, which helps with the stability and accuracy during model training; Use polynomial feature generation or feature crossing methods to capture the non-linear relationships between features. This helps to discover the complex associations hidden between user behavior and transaction features; Use deep learning techniques (such as embedding layers) to perform dimensionality reduction and representation learning on high-dimensional sparse features, improving the effectiveness of features and the generalization ability of the model; Use the K-means algorithm to perform clustering analysis on user behavior data to identify user groups with similar behavior patterns. Through visual display (such as two-dimensional or three-dimensional scatter plots after t-SNE or PCA dimensionality reduction), it helps business personnel intuitively understand the distribution and differences of user behavior; Perform clustering analysis on transaction records to identify user groups or product categories with similar transaction features. Combining the clustering results with business knowledge, interpret and analyze the cluster centers to extract valuable insights, such as potential market segmentation or user portraits; Use association rule mining algorithms (such as Apriori, FP-Growth, etc.) to mine the association relationships between products from transaction data. This helps to discover the patterns of user purchase behavior, such as "users who buy milk often also buy bread"; Combining user behavior data and transaction records, mine the association relationships between user behaviors. This helps to understand the conversion paths between different behavior stages of users, such as "users who have browsed a certain type of product are more likely to buy another type of product"; Build user behavior sequence models (such as hidden Markov models, recurrent neural networks, etc.) to capture the patterns and trends in user behavior sequences. This helps to predict users' future behaviors, such as the products users may buy or the pages they may visit; By analyzing user behavior sequences, we can deeply understand users' purchase habits, preferences, and potential needs, providing a scientific basis for personalized recommendations and marketing strategies.

[0112] The effects of the above technical solution are as follows: By extracting multi-dimensional features from user behavior data, transaction records, system logs, and third-party data sources, the comprehensiveness and accuracy of the data are ensured. The application of the feature selection algorithm further filters out the features that have a significant impact on the analysis target, reducing the interference of noise data; encoding and normalizing categorical features and numerical features makes the data format unified, facilitating the processing and analysis of subsequent algorithms. This helps to improve the automation degree and efficiency of the analysis process; by using polynomial feature generation or feature crossing methods, the non-linear relationships between features are captured, which helps to discover complex patterns hidden in the data and enhance the prediction ability of the model; using deep learning techniques to perform dimensionality reduction and representation learning on high-dimensional sparse features improves the effectiveness of the features and the generalization ability of the model, helping to obtain better analysis results with limited data; by performing clustering analysis on user behavior data using the K-means algorithm, user groups with similar behavior patterns are identified, providing an important basis for constructing user portraits. This helps enterprises better understand user needs and behavior habits; performing clustering analysis on transaction records to identify user groups or product categories with similar transaction characteristics provides strong support for market segmentation. This helps enterprises formulate more accurate marketing strategies and product positioning; by using the association rule mining algorithm to mine the association relationships between products from transaction data, the laws of user purchase behavior are revealed. This helps enterprises optimize product portfolios and promotion strategies to increase sales; by combining user behavior data and transaction records, mining the association relationships between user behaviors provides enterprises with deeper insights into user behaviors. This helps enterprises predict users' future purchase trends and develop personalized recommendation algorithms; by constructing a user behavior sequence model, the patterns and trends in the user behavior sequence are captured. This helps enterprises more accurately predict users' future behavior paths and purchase intentions; based on the analysis results of the user behavior sequence model, the recommendation algorithm can be optimized to improve the accuracy of recommendations and user satisfaction. This helps to enhance the user experience and loyalty; through comprehensive and in-depth user behavior analysis and understanding of transaction characteristics, more scientific and accurate data support is provided for enterprises. This helps enterprises make more reasonable business decisions and strategic plans; by deeply mining and analyzing user behaviors and transaction characteristics, enterprises can discover new market opportunities and innovation points, promoting the continuous growth and innovative development of the business.

[0113] An embodiment of the present invention, as Figure 2 shown, is a three-party transaction intelligent feedback optimization system, and the system includes:

[0114] Data Acquisition Module: Real-time collect transaction data from multiple data sources (such as e-commerce platforms, financial services, and logistics, etc.) through a distributed data acquisition algorithm. The transaction data includes transaction records, user behavior logs, product information, and payment details; and preprocess the collected transaction data.

[0115] Feature Extraction Module: Apply a deep learning model to perform deep feature extraction on the preprocessed data to capture complex features such as user behavior and product attributes; Use a graph neural network to construct a transaction relationship graph, with entities as nodes, where the entities include users, products, and transactions, and transaction behaviors as edges to form a complex transaction network.

[0116] Behavior Detection Module: Based on deep features, use a deep belief network for user sentiment analysis to identify positive, negative, or neutral sentiment in user feedback; Combine graph neural network and reinforcement learning techniques to construct a dynamic fraud detection model, and through simulating the transaction process, real-time detect potential fraud behaviors such as false transactions and money laundering.

[0117] Strategy Adjustment Module: Through a deep learning recommendation system (such as deep collaborative filtering, neural collaborative filtering), combine user historical behavior, real-time feedback, and the transaction relationship graph to generate personalized product, service, or discount recommendations; According to the feedback of transaction data, apply deep reinforcement learning to dynamically adjust transaction strategies such as price adjustment and inventory allocation.

[0118] Process Recording Module: Based on an intelligent feedback loop mechanism, collect and analyze user feedback in real-time, continuously iterate and optimize recommendations and transaction strategies, and use blockchain technology to record the process of each feedback analysis and strategy adjustment.

[0119] The working principle of the above technical solution is as follows: collect data in real time from multiple reliable data sources to ensure the diversity and comprehensiveness of the data. The data sources may include e-commerce platforms, financial service providers, logistics companies, etc.; adopt a distributed data collection algorithm to ensure data synchronization between different data sources and avoid data latency or inconsistency; remove duplicate, invalid or incorrect data, such as null values, outliers, etc. At the same time, perform deduplication and denoising processing on the data to improve data quality; format the data from different sources into a unified format for subsequent processing and analysis. This may include data type conversion, data structure adjustment, etc.; perform standardization or normalization processing on data of different dimensions to ensure their comparability in the model. This helps to reduce biases and errors in the model training process; select an appropriate deep learning model for feature extraction according to the characteristics of the data and business requirements. Commonly used models include convolutional neural networks (CNNs), recurrent neural networks (RNNs), etc.; extract valuable features for subsequent analysis from the original data, such as user behavior features, product attribute features, etc. These features can reflect information such as user preferences and market trends; fuse features from different sources to form a more comprehensive and rich feature set. This helps to improve the accuracy and robustness of the model; use users, products, and transactions as nodes in the graph. Each node has a unique identifier and attribute information; use transaction behaviors as edges to connect user, product, and transaction nodes. The weights of the edges can represent information such as transaction amount and transaction frequency; use graph neural networks to learn the features of nodes and edges and capture potential relationships in the transaction network. This helps to discover social relationships between users, similarities between products, etc.; use a deep belief network to perform sentiment analysis on user feedback. By training the model, identify positive, negative, or neutral sentiments in user feedback; generate sentiment labels for each piece of user feedback for subsequent analysis and recommendation. This helps to understand users' satisfaction with transactions and products and provides a basis for improving services; combine graph neural networks and reinforcement learning techniques to build a dynamic fraud detection model. The model can simulate the transaction process and detect potential fraud behaviors in real time; through training the model, identify fraud behaviors such as false transactions and money laundering. At the same time, the model can continuously optimize the fraud detection strategy according to the detection results to improve the accuracy and efficiency of detection; use a deep learning recommendation system to generate personalized recommendations by combining user historical behaviors, real-time feedback, and transaction relationship graphs. The recommendation system can consider factors such as user preferences, market trends, and similarities between products to provide accurate recommendations; continuously adjust the recommendation strategy according to users' feedback and transaction data to optimize the recommendation results. This helps to improve users' satisfaction and loyalty; apply deep reinforcement learning to dynamically adjust transaction strategies, such as price adjustment, inventory allocation, etc. The model can continuously optimize the strategy according to transaction data feedback to improve transaction efficiency and profitability; apply the adjusted transaction strategy to the actual business and conduct real-time monitoring and evaluation.Adjust the strategy in a timely manner according to the execution effect of the strategy to ensure the steady development of the business; collect user feedback in real time based on the intelligent feedback loop mechanism. This includes information such as user evaluations, suggestions, etc. on transactions, goods, and services; analyze and process the collected feedback, and extract valuable information for optimizing recommendation and trading strategies. This helps to continuously improve the business and enhance user satisfaction; use blockchain technology to record the process of each feedback analysis and strategy adjustment. This includes information on all aspects such as feedback collection, processing, analysis, and optimization; the blockchain technology has the characteristic of data immutability, which can ensure the accuracy and reliability of the records. This provides strong support for subsequent audits and traceability.

[0120] The effects of the above technical solutions are as follows: through automated and intelligent data collection, processing, and analysis processes, manual intervention and delays are reduced; the application of deep learning models and graph neural networks improves the accuracy of feature extraction and transaction relationship recognition; the personalized recommendation system can generate recommendation content that meets user needs based on the user's historical behavior and real-time feedback; the dynamically adjusted trading strategy can better meet the changes in market demand and user needs; the fraud detection model that combines deep belief networks and graph neural networks with reinforcement learning can detect potential fraud behaviors in real time; the application of blockchain technology ensures the transparency and traceability of the transaction process and improves the security of transactions; the intelligent feedback loop mechanism can collect and analyze user feedback in real time; continuously iterate and optimize recommendation and trading strategies to adapt to the changing market environment and user needs.

[0121] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A three-party transaction intelligent feedback optimization method, characterized in that: The method comprises: S1. Use distributed data collection algorithms to collect transaction data from multiple data sources in real time and pre-process the collected transaction data; S2. Apply deep learning models to extract deep features from preprocessed data, and use graph neural networks to construct transaction relationship graphs, taking entities as nodes and transaction behaviors as edges to form a complex transaction network. S3. Based on deep features, a deep belief network is used to perform user sentiment analysis to identify positive, negative or neutral sentiment in user feedback; a dynamic fraud detection model is built by combining graph neural networks and reinforcement learning technology to detect potential fraudulent behavior in real time by simulating transaction processes; S4. Generate personalized product, service or discount recommendations by combining user historical behavior, real-time feedback and transaction relationship graph through deep learning recommendation system; dynamically adjust transaction strategies based on transaction data feedback using deep reinforcement learning; S5. Based on the intelligent feedback loop mechanism, user feedback is collected and analyzed in real time, and recommendation and trading strategies are continuously optimized and iterated. Blockchain technology is used to record the process of each feedback analysis and strategy adjustment.

2. According to claim 1, a three-party transaction intelligent feedback optimization method is characterized in that: Said S1 comprises: S11. Identify and access multiple key data sources, including e-commerce platforms, financial service systems, and logistics information systems; use API interfaces to access data in real time and stably; S12. Determine the type of transaction data, which includes transaction records, user behavior logs, product information, and payment details; S13. Classify and store the data according to data type; and pre-process the stored data.

3. According to claim 1, a three-party transaction intelligent feedback optimization method is characterized in that: The S2 comprises: S21. Perform feature selection based on business requirements and data characteristics, and preliminarily extract features through traditional feature extraction methods to form a preliminary feature set; S22. Based on the initially extracted features and data characteristics, feature extraction is performed based on the deep learning model, and the architecture of the deep learning model is designed; S23, training the model, using the trained deep learning model to perform deep feature extraction on the initially extracted features, digging out deeper feature information, integrating the deep features with traditional features to form a comprehensive feature set; S24. Perform advanced feature engineering on the fused features; use feature importance evaluation methods to screen out key features that have a significant impact on model performance; and further process and optimize the key features; S25. Construct a transaction relationship graph based on the entities in the transaction data and the interaction relationships between them; use the entities as nodes and the transaction behaviors as edges to form a complex transaction network; S26. Preprocess the transaction relationship graph; learn and represent it through a graph neural network model according to the characteristics and requirements of the transaction relationship graph.

4. According to claim 3, a three-party transaction intelligent feedback optimization method is characterized in that: The S23 comprises: Based on the selected deep learning model, a large-scale data set is used for model training. The trained deep learning model is used to perform multi-level and multi-scale deep feature extraction on the initially extracted features. Through the hidden layer, feature information at different levels is mined; and the extracted deep features are preprocessed; Based on the model’s feature importance evaluation, deep features that have an impact on subsequent tasks are selected; deep features are fused with traditional features through nonlinear combination; Further optimize the fused features and screen them again through feature selection based on statistics; Use the validation data set to validate the fused features and evaluate their performance in subsequent tasks; iteratively adjust the feature extraction, fusion strategy, and optimization process based on the validation results; Perform multimodal fusion of features from different modalities; use multimodal deep learning models to extract and fuse cross-modal features.

5. According to claim 3, a three-party transaction intelligent feedback optimization method is characterized in that: The S25 comprises: Identify key entities from transaction data and classify entities based on business needs and data characteristics; use natural language processing technology to extract entities from text data; Design attributes for each entity node and add timestamps to the node attributes; identify the interaction between users and products from transaction data and classify them according to the importance and frequency of the behavior; Through machine learning algorithms, transaction behaviors are automatically classified and labeled, attributes are designed for each edge, and additional dimensions are added to edge attributes; Connect the identified entity nodes and transaction behavior edges according to business logic and data characteristics to build a preliminary transaction relationship graph; Add global attributes to the transaction relationship graph and convert the graph structure into a low-dimensional vector representation through graph embedding technology; use graph algorithms to evaluate the importance of nodes and edges and remove noise nodes and edges; According to the properties of nodes and edges, a weighted algorithm is used to calculate the weights of nodes and edges, and the weights of nodes and edges are dynamically adjusted based on time factors; Introduce the time dimension, build a dynamic transaction relationship diagram, use time series analysis technology to perform time series prediction and trend analysis on the transaction relationship diagram, and adjust and optimize the transaction relationship diagram based on the analysis results.

6. According to claim 1, a three-party transaction intelligent feedback optimization method is characterized in that: The S3 includes: S31. Perform sentiment analysis on user feedback through deep belief networks to identify positive, negative or neutral sentiments; and use the sentiment analysis results to evaluate user satisfaction; S32. Combine graph neural network and reinforcement learning technology to build a dynamic fraud detection model and set up fraud detection strategies; S33, using a reinforcement learning algorithm to train the dynamic fraud detection model and perform real-time detection of potential fraudulent behaviors; S34. Mark and record the identified fraudulent activities and take appropriate response measures based on the nature and severity of the fraudulent activities.

7. According to claim 1, a three-party transaction intelligent feedback optimization method is characterized in that: The S4 comprises: S41. Use fusion algorithms to fuse user historical behavior data, real-time feedback data, and transaction relationship graph data to form a comprehensive user portrait; S42. Perform feature engineering and feature extraction on the fused data; build a recommendation system through a deep learning model; and design a variety of recommendation strategies; S43. Evaluate and optimize the recommended strategy through A / B testing and select the best recommended strategy; conduct in-depth analysis of trading data and extract indicators that have a significant impact on the adjustment of trading strategies; S44. Use data mining technology to explore potential rules and patterns in transaction data, and build a deep reinforcement learning model based on the results of transaction data analysis; use historical transaction data to train the model, and perform continuous iteration and optimization; S45. Apply the decision results of the DRL model to the trading strategy adjustment algorithm to perform intelligent and automated adjustment of the trading strategy; S46. Evaluate the effectiveness of the adjusted trading strategy by comparing the indicators before and after the adjustment; feed back the evaluation results to the DRL model to guide further optimization and training of the model.

8. According to claim 7, a three-party transaction intelligent feedback optimization method is characterized in that: The S44 comprises: Through the association rule mining algorithm, the association relationship between commodities is mined from the transaction data, and the mined association rules are visualized; Use clustering algorithms to cluster users or products, interpret and analyze the clustering results, and extract the characteristics of the cluster centers; Through the time series analysis algorithm, the transaction data is predicted in time series to predict future sales trends and user behavior changes; combined with relevant factors, the time series prediction results are corrected and optimized; According to the needs of trading strategy adjustment and data characteristics, build the architecture of deep reinforcement learning model, define state space, and design action space; Design reward functions according to business objectives, weight and combine reward functions, use historical transaction data for model training, and update model parameters through iterative optimization algorithms; The model is trained and tested through the simulation system, and the model architecture, state space, action space and reward function are adjusted and optimized based on the training results and test feedback.

9. According to claim 1, a three-party transaction intelligent feedback optimization method is characterized in that: The S5 comprises: S51, integrating feedback data from multiple channels and preprocessing the collected feedback data; S52. Based on machine learning or deep learning technology, extract key features from feedback data, encode and transform them, analyze user behavior patterns and transaction characteristics through clustering and association rule mining, and identify potential market trends and user needs; S53. Use natural language processing technology to conduct sentiment analysis on text data to understand users’ satisfaction and opinions on products and services; S54. Based on historical data and features, build a prediction model to predict users' future behavior trends and transaction intentions, and instantly capture and process user feedback through a real-time feedback system; continuously adjust and optimize the collection, processing, analysis and application of feedback data through iterative optimization of feedback loops; S55. Intelligently adjust the recommendation algorithm and transaction strategy based on the feedback analysis results, select the blockchain consensus mechanism based on business needs and data characteristics, and deploy the blockchain network; S56. Put key information on the chain through smart contracts, use the immutability of blockchain to verify and audit data, and based on the data access control mechanism, only authorized users or systems can access and modify data on the blockchain.

10. A three-party transaction intelligent feedback optimization system, characterized in that: The system comprises: Data collection module: collects transaction data from multiple data sources in real time through distributed data collection algorithms, and pre-processes the collected transaction data; Feature extraction module: Apply deep learning models to extract deep features from preprocessed data, use graph neural networks to build transaction relationship graphs, use entities as nodes and transaction behaviors as edges, and form a complex transaction network; Behavior detection module: Based on deep features, deep belief networks are used to analyze user sentiment and identify positive, negative or neutral sentiment in user feedback. A dynamic fraud detection model is built by combining graph neural networks and reinforcement learning technology to detect potential fraudulent behavior in real time by simulating transaction processes. Strategy adjustment module: Generate personalized product, service or discount recommendations by combining user historical behavior, real-time feedback and transaction relationship graph through deep learning recommendation system; dynamically adjust transaction strategies based on transaction data feedback using deep reinforcement learning; Process recording module: Based on the intelligent feedback loop mechanism, it collects and analyzes user feedback in real time, continuously iterates and optimizes recommendation and trading strategies, and uses blockchain technology to record the process of each feedback analysis and strategy adjustment.

Images