Network transaction credit risk intelligent evaluation system and method based on multi-modal data fusion and deep learning

Abstract

The invention discloses an intelligent network transaction credit risk assessment system and method based on multi-modal data fusion and deep learning, and particularly relates to the technical field of network transaction risk assessment, and the system comprises a multi-modal data collection, fusion and deep reinforcement learning agent, and a large model interaction and risk assessment and supervision execution module. The multi-modal data acquisition module acquires various data in real time; the fusion module realizes semantic alignment and interactive learning through multiple strategies; the deep reinforcement learning agent module optimizes a monitoring and supervision strategy based on an Actor-Critic framework; the large model interaction module provides multi-aspect support by means of various technologies; and the risk assessment and supervision execution module identifies illegal behaviors and executes supervision actions. The evaluation method comprises the steps of data acquisition and preprocessing, feature vector generation, agent training, large model deployment optimization, real-time monitoring decision and the like. The system can improve the monitoring and supervision efficiency, protects the rights and interests of consumers, constructs a healthy transaction environment, and promotes the development of related technologies.

Description

Technical Field

[0001] The present invention relates to the technical field of network transaction risk assessment. More specifically, the present invention relates to an intelligent assessment system and method for network transaction credit risk based on multi-modal data fusion and deep learning. Background Art

[0002] Data from the data center of the National E-commerce Public Service Network shows that the e-commerce transaction volume reached 43.83 trillion in 2022, of which the national online retail sales reached 13.7 trillion; as of June 2023, there were more than 70 million live-streamed products on key monitored e-commerce platforms, with cumulative sales of 1.27 trillion yuan, and the number of active live-stream hosts reached more than 2.7 million. As of June 2024, there were a total of 79 e-commerce platforms in a certain province, and the transaction amount achieved on the platforms in this province was 254.84 billion yuan, a year-on-year increase of 29.0%. At the same time, there are many problems in network transactions, mainly manifested as: problems with the identity of business entities and violations of business qualifications, violations of transaction behaviors such as false publicity, price violations, brushing orders and boosting credit, and violations of live-stream content, violations of transaction commodities such as product quality problems, sales of prohibited items, and infringement of intellectual property rights, and violations of transaction platform entities such as unfair contract terms and leakage of personal information and privacy. The technical research on the monitoring and supervision of network transaction data mainly focuses on fields such as web crawler technology, web page embedding technology, network probe technology, and database probe technology.

[0003] With the explosion of commodity types on network transaction platforms and the diversification of user interaction methods, a single data modality can no longer meet the need for a comprehensive understanding of commodities. Through effective data fusion technology, artificial intelligence models can more accurately capture the internal connections between multi-modal data, thus playing a key role in monitoring and supervision behaviors, market trend analysis, etc.

[0004] However, when the existing technology solves the above problems in network transactions, there are still the following drawbacks:

[0005] First, network transaction data is complex and diverse, making it difficult to analyze effectively. Network transaction data covers multi-modal data such as text, images, videos, and audio, and the data volume is huge, with different formats and structures. Traditional manual analysis methods are difficult to effectively process the massive data and difficult to extract valuable information from multi-modal data.

[0006] Second, network transaction behaviors are complex and changeable, making it difficult to identify effectively. There are many types of network transaction behaviors and they are constantly changing, such as false publicity, brushing orders and boosting credit, price fraud, infringement of intellectual property rights, etc.; traditional manual identification methods are difficult to accurately identify complex violations and are inefficient.

[0007] Third, there are technical bottlenecks, lacking effective supervision tools and methods.

[0008] Therefore, an intelligent evaluation system and method for network transaction credit risk based on multi-modal data fusion and deep learning are proposed Summary of the Invention

[0009] To overcome the above defects of the prior art, the present invention provides an intelligent evaluation system and method for network transaction credit risk based on multi-modal data fusion and deep learning to solve the problems raised in the above background technology

[0010] To achieve the above object, the present invention provides the following technical solutions: An intelligent evaluation system for network transaction credit risk based on multi-modal data fusion and deep learning, including:

[0011] A multi-modal data acquisition module for real-time acquisition of text, images, videos, audio, and user behavior data of the network transaction platform

[0012] A multi-modal data fusion module that performs semantic alignment and cross-modal interactive learning on multi-modal data through feature-level fusion, model-level fusion, and decision-level fusion strategies to generate a unified multi-modal feature vector

[0013] A deep reinforcement learning agent module constructed based on the Actor-Critic network framework, which optimizes the monitoring and supervision strategy through dynamic data interacting with the environment, including a Critic network for evaluating state-action value and an Actor network for generating optimal supervision actions

[0014] A large model interaction module that integrates a pre-trained large model and provides multi-modal feature extraction, virtual environment simulation, and auxiliary decision-making support through prompting engineering, retrieval-augmented generation, and supervised fine-tuning techniques

[0015] A risk assessment and supervision execution module that identifies violations according to the strategy output by the agent and generates a risk warning signal, and executes differentiated supervision actions

[0016] Preferably, the multi-modal data fusion module uses the following methods to achieve data alignment:

[0017] Capture the semantic association between text and images through cross-modal attention mechanism

[0018] Use masked self-supervised learning to impute missing modal data

[0019] Based on multi-modal embedding technology, achieve entity alignment and sentiment polarity matching

[0020] Preferably, in the deep reinforcement learning agent module:

[0021] The Critic network uses a deep Q-network (DQN) combined with temporal difference learning (TDL) to optimize the value function

[0022] The Actor network maximizes the long-term cumulative reward through the policy gradient method and introduces a hierarchical task decomposition mechanism;

[0023] The experience pool uses the priority replay technique to store interaction data, enhancing the diversity of training samples and the stability of the policy.

[0024] Preferably, the large model interaction module includes:

[0025] A vision-language pre-trained model based on Kaleido-BERT extracts the joint features of product text and images;

[0026] Utilize the large model to generate simulated transaction data and user behavior sequences, constructing a reinforcement learning training environment;

[0027] Combine real-time monitoring data with historical risk patterns through the chain of thought technique to generate interpretable decision-making suggestions.

[0028] Preferably, the risk assessment and regulatory enforcement module includes:

[0029] Identify false propaganda, brushing and boosting credit, and the sale of prohibited items through multi-modal feature matching;

[0030] Detect big data price discrimination and price violation behaviors based on the analysis of user behavior sequences;

[0031] Dynamically adjust the regulatory policy threshold to achieve the balanced optimization of the compliance rate and the recall rate of violation identification.

[0032] An evaluation method for the network transaction credit risk intelligent evaluation system based on multi-modal data fusion and deep learning as described above, comprising the following steps:

[0033] S1: Collect multi-modal data from the network transaction platform and perform cleaning, format unification, and time series alignment;

[0034] S2: Generate a multi-modal joint feature vector through feature-level fusion and cross-modal attention mechanism;

[0035] S3: Construct a deep reinforcement learning agent and train the Actor-Critic network using simulated environment data and real interaction data;

[0036] S4: Deploy the pre-trained large model and optimize the risk identification rules through prompt engineering and retrieval-augmented generation techniques;

[0037] S5: Monitor transaction behaviors in real time and output the risk level and regulatory instructions by combining the agent's policy and the large model's auxiliary decision-making.

[0038] Preferably, the training process in step S3 includes:

[0039] Design a dual reward mechanism to provide positive incentives for compliance and impose penalties on violations;

[0040] The target network delayed update strategy is used to stabilize the parameter optimization of the critic network;

[0041] Respond to dynamic changes in trading environment through adaptive learning rate adjustment module.

[0042] Preferably, in step S4:

[0043] Use the CommerceMM model to perform end-to-end fusion of multimodal e-commerce data;

[0044] Retrieve historical regulatory case libraries based on RAG technology to enhance the generation of risk response strategies;

[0045] Infuse domain knowledge into large model parameters through supervised fine-tuning techniques.

[0046] Preferably, the step S5 further includes:

[0047] Establish an interpretable analysis module to visualize the basis of risk decision-making through attention weights;

[0048] It supports supervisors to interactively adjust the strategy parameters of intelligent agents to achieve human-machine collaborative optimization.

[0049] Technical effects and advantages of the present invention:

[0050] Through deep reinforcement learning, monitoring and regulatory intelligent agents can analyze online transaction data in real time, accurately identify potential risks, and automatically formulate and implement differentiated monitoring and regulatory strategies. This will effectively improve the monitoring and regulatory efficiency of government departments and platform companies, reduce labor costs, and ensure the healthy and orderly development of the online transaction market.

[0051] By intelligently identifying violations such as false advertising, counterfeiting, fake orders, and big data price discrimination, deep reinforcement learning monitoring and supervision agents can effectively curb the fraudulent behavior of illegal businesses, protect the legitimate rights and interests of consumers, enhance consumers' trust in online transactions, promote the growth of transaction volume, and thereby expand the scale of the online transaction market.

[0052] The present invention is conducive to building a fairer, more transparent and secure online transaction environment, effectively preventing unfair competition and market monopoly, and providing strong guarantees for the sustained and healthy development of the digital economy. At the same time, it applies cutting-edge technologies such as deep reinforcement learning and large models to the field of online transaction monitoring and supervision, which not only enriches the scenarios of technology application, but also promotes the innovation and development of related technologies, and is conducive to driving the upgrading and transformation of related industrial chains such as artificial intelligence, vertical large models, offline logistics, and online payments, and promoting the healthy development of related enterprises in the field of online transactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a system technology framework diagram of the present invention.

[0054] Figure 2 It is a schematic diagram of the operation method of the multimodal data fusion module in the network transaction field of the present invention.

[0055] Figure 3 Schematic diagram of the deep reinforcement learning module operation method of the present invention.

[0056] Figure 4 It is a schematic diagram of the operation method of the large model interaction module of the present invention. DETAILED DESCRIPTION

[0057] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.

[0058] As attached Figures 1-4 The network transaction credit risk intelligent assessment system based on multimodal data fusion and deep learning shown includes:

[0059] Multimodal data collection module, used to collect text, images, videos, audio and user behavior data of online trading platforms in real time;

[0060] In specific implementation, the data sources come from the stock and incremental multimodal data provided by online trading platforms (large e-commerce platform companies), customized data sets from third-party institutions, and publicly published data sets, covering multiple modalities such as text, images, videos, and audios. The content includes but is not limited to basic information of platforms and online stores, user comments, product image libraries, audio and video, transaction data, news and public opinion, complaint information, etc. To ensure the accuracy and applicability of the data set, the system performs preprocessing, involving steps such as data cleaning, format unification, and sequence alignment.

[0061] Feature extraction, as another core part of the preprocessing, is implemented using a pre-trained large model. At the same time, for some texts and images, natural language processing techniques such as Long Short-Term Memory (LSTM) are planned to be used to identify keywords and sentiment polarity in the text; computer vision techniques such as Residual Network (ResNet) are used for the recognition of objects and scenes in images, so as to provide richer information representations for multi-modal data fusion.

[0062] As shown in the Figure 2 appendix, the multi-modal data fusion module performs semantic alignment and cross-modal interactive learning on multi-modal data through feature-level fusion, model-level fusion, and decision-level fusion strategies, generates a unified multi-modal feature vector, and constructs a multi-modal data fusion model;

[0063] The multi-modal data fusion module realizes data alignment by the following methods:

[0064] Capture the semantic association between text and image through a cross-modal attention mechanism;

[0065] Specifically, calculate the cross-modal association weight through the attention mechanism, and the formula is:

[0066]

[0067] where Q text and K image are the query matrix and key matrix of text and image respectively, and d is the dimension;

[0068] Use masked self-supervised learning to impute missing modal data; among them, the missing function is:

[0069]

[0070] where M is the set of missing modalities, is the modal data after imputation.

[0071] Based on multi-modal embedding technology, entity alignment and sentiment polarity matching are realized.

[0072] Specifically, feature-level fusion: Concatenate the feature vectors of each modality into F concat =[F txt ; F image ;

[0073] Model-level fusion: Capture cross-modal associations through the Transformer encoder to generate a joint representation, and M = LzyerNorm(MultiHeadAttention(X,Y)+FFN(X))

[0074] Among them, X and Y are inputs of different modalities, and FFN is a feed-forward network;

[0075] Decision-level fusion: Weighted voting is performed on the outputs of each modality independent model, and the formula is:

[0076]

[0077] Finally, a multi-task learning framework is adopted to jointly optimize the separation loss and the alignment loss and through multi-dimensional evaluation metrics such as accuracy, recall, and F1 score, the optimization and iteration of the model are realized to meet the requirements of actual monitoring and supervision application scenarios.

[0078] As shown Figure 3 in the appendix, the deep reinforcement learning agent module is constructed based on the Actor-Critic network framework, and the monitoring and supervision strategy is optimized through the dynamic data of interacting with the environment, including the Critic network for evaluating the state-action value and the Actor network for generating the optimal supervision action;

[0079] In the deep reinforcement learning agent module:

[0080] The Critic network uses the deep Q-network (DQN) combined with temporal difference learning (TDL) to optimize the value function;

[0081] Specifically, the agent selects the action a t according to the current state s t , receives the reward r t and enters the new state s t+1 , and the data is stored in the experience pool. Among them, the target Q value is calculated as:

[0082]

[0083] and the Q value function is updated by minimizing the Bellman error:

[0084]

[0085] Among them, θ' and are the target network parameters, and γ is the discount factor.

[0086] The Actor network maximizes the long-term cumulative reward through the policy gradient method and introduces a hierarchical task decomposition mechanism;

[0087] The policy parameters are updated through the policy gradient, and the gradient formula is:

[0088]

[0089] The experience pool uses priority replay technology to store interaction data to improve the diversity of training samples and the stability of strategies. The sampling probability is:

[0090]

[0091] p i =|δ i |+∈

[0092] Among them, δ i is the timing difference error, α is the priority adjustment coefficient;

[0093] In the specific implementation, (1) environmental interaction and data collection. The "model-free reinforcement learning" technology based on data interaction is mainly adopted. There is no need to set a specific static model. The dynamic learning strategy is learned through the continuous input of interactive user data and monitoring and supervision data. At the same time, a dual reward mechanism is proposed. On the one hand, it rewards online trading platforms for adopting compliance strategies, such as complying with regulatory rules and protecting user privacy; on the other hand, it punishes online trading platforms for illegal behaviors, such as false propaganda and big data killing old customers. In the interaction, the intelligent agent is based on the current state s t , environmental feedback reward value r t Select action a t , thus entering the next state s t+1 The entire interaction data is recorded in real time and stored in the experience pool. By acting on the experience pool on the Critic network and the Actor network, the randomness and diversity of data samples are ensured, and the stability and efficiency of subsequent learning are improved.

[0094] (2) Critic network training. After accumulating enough interaction data, we plan to adopt technologies such as temporal-difference learning (TDL) or deep Q-network (DQN) to build a critic network. At the same time, we use the data in the experience pool to train the critic network and update the network parameters by minimizing the difference between the value estimate and the actual return. The training of the critic network provides a value benchmark for the actor network update, namely the advantage function, which is used to measure the advantage of executing a decision relative to the average situation under a specific state.

[0095] (3) Actor network update. This mainly uses the interaction data of multimodal fusion in the field of online transactions and the evaluation results of the Critic network to train the Actor network around the goal of maximizing long-term cumulative rewards, so that it can select the optimal strategy that meets the monitoring and supervision rules according to the needs of online transaction monitoring and supervision. This study intends to achieve this through the policy gradient method, in which the parameter update direction of the Actor network is proportional to the gradient of the advantage function, ensuring that the behavior of the intelligent agent gradually tends to the optimal strategy.

[0096] (4) Model Iteration and Optimization. Specifically, the agent updates and iterates the model based on its overall performance and evaluation results under the current policy. Optimization algorithms, experience replay, exploration-exploitation strategies, policy gradient methods, batch normalization, regularization techniques, parameter initialization techniques, and standardized batch processing techniques in machine learning are used to optimize the model performance. The performance of the entire model (e.g., indicators such as the compliance rate of the online trading platform and the credit evaluation of online stores) will be evaluated using test data during model iteration, improving the learning efficiency of the agent and enhancing its adaptability and robustness in different environments.

[0097] As shown in the Figure 4 appendix, the large model interaction module integrates a pre-trained large model and provides multi-modal feature extraction, virtual environment simulation, and auxiliary decision support through prompting engineering, retrieval-augmented generation, and supervised fine-tuning techniques;

[0098] The large model interaction module includes:

[0099] Extract the joint features of commodity text and images using a vision-language pre-trained model based on Kaleido-BERT;

[0100] Generate simulated trading data and user behavior sequences using the large model to construct a reinforcement learning training environment;

[0101] Combine real-time monitoring data with historical risk patterns through the chain of thought technique to generate interpretable decision-making suggestions.

[0102] During specific implementation, interactions are carried out using the large model in aspects such as pre-training, feature representation, output of specific layers, and feature adjustment.

[0103] In terms of pre-training, a general pre-trained large model is used to process unlabeled data, including training text data using language modeling tasks (such as autoregressive tasks or masked language modeling) and training visual data using image classification or image-text matching tasks, thereby capturing general features and patterns in the data.

[0104] In terms of feature representation, the pre-trained large model outputs the embedding vectors of each token in the text and the feature vectors at the pixel or region level in the image.

[0105] In terms of the output of specific layers, for tasks that require fine-grained features, lower layer outputs are selected; for tasks that require more abstract features, higher layer outputs are selected.

[0106] In terms of feature adjustment, according to the data fusion situation and the monitoring and regulatory environment, the most relevant part of the features is selected from the model output features, or the feature quality is improved through methods such as feature scaling and normalization.

[0107] Based on real online transaction data, use large models to expand and enhance the data, generating simulated data (multi-modal data), simulated user behaviors (such as transaction behavior data), and simulated regulatory strategies (such as regulatory strategies for big data price discrimination), so as to interact with the deep reinforcement learning A-C network model.

[0108] In terms of exploration and exploitation, use large models to simulate the external online transaction environment and regulatory strategies, and through reinforcement learning research, make a balance between exploration (trying new actions) and exploitation (using the known best strategies).

[0109] In terms of shared representation, use large models to help the A-C network process high-dimensional inputs (such as images or texts) and understand complex state representations; in terms of parameter sharing, use large models to reduce the number of parameters, reduce the risk of overfitting, and improve the generalization ability of the model.

[0110] In terms of model training, use large models to update the Actor network and the Critic network simultaneously in a unified training process to improve the training effect; or in some variant algorithms of the A-C network, introduce a target network to stabilize the training process.

[0111] The risk assessment and regulatory execution module identifies violations according to the strategies output by the agent and generates risk warning signals, and executes differentiated regulatory actions;

[0112] The risk assessment and regulatory execution module includes:

[0113] Identify false propaganda, brush trading volume and hype, and sales of prohibited items through multi-modal feature matching;

[0114] Detect big data price discrimination and price violation behaviors based on user behavior sequence analysis;

[0115] Dynamically adjust the regulatory strategy threshold to achieve balanced optimization of the compliance rate and the recall rate of violation identification.

[0116] In specific implementation, by designing a workflow, call different large models to divide risk identification into sub-tasks such as anomaly monitoring, false propaganda, counterfeits, brush trading volume and hype, and big data price discrimination, and guide the operations of monitoring and alarm instructions through specific instruction templates; by designing a chain of thought, store the observed current state in short-term memory for quickly processing real-time data streams, and then combine it with historical patterns in long-term memory to infer potential risks through inference learning.

[0117] At the application level, big models are used to conduct risk assessments on transactions, helping regulators identify high-risk transactions and monitor transaction activities in real time; big models are used to understand information in product descriptions, images, and user comments, helping to identify counterfeit and substandard products or illegal information; big models are used to analyze user historical behavior, identify normal and abnormal behavior patterns, and analyze the behavior of specific user groups to issue public opinion warnings; big models are used to interact with decision makers to answer immediate questions about data or predictions, and the big model explanation function is used to help decision makers understand why a specific decision path is recommended.

[0118] The deployment of large models usually requires techniques such as prompt engineering, retrieval-enhanced generation, and supervised fine-tuning. Prompt engineering technology guides the large model to generate the expected output by designing instruction steps, so that the intelligent agent can obtain network transaction risk assessment and monitoring and supervision suggestions from the large model; retrieval-enhanced generation technology mainly retrieves monitoring and supervision information related to the current task, combines previous knowledge and experience, and improves the accuracy and diversity of decision-making; supervised fine-tuning is to further train the large model in a specific monitoring and supervision scenario, so that the intelligent agent can learn more fine-grained monitoring and supervision rules and strategies, thereby achieving more accurate and efficient operations in risk identification and differentiated monitoring and supervision tasks.

[0119] An evaluation method for an intelligent evaluation system for online transaction credit risk based on multimodal data fusion and deep learning, comprising the following steps:

[0120] S1: Collect multimodal data from online trading platforms, clean them, unify the format and align the time series.

[0121] S2: Generate multimodal joint feature vector through feature level fusion and cross-modal attention mechanism.

[0122] S3: Build a deep reinforcement learning agent and train the Actor-Critic network using simulated environment data and real interaction data;

[0123] The training process in step S3 includes:

[0124] Design a dual reward mechanism to provide positive incentives for compliance and impose penalties on violations;

[0125] The target network delayed update strategy is used to stabilize the parameter optimization of the critic network;

[0126] Respond to dynamic changes in trading environment through adaptive learning rate adjustment module.

[0127] S4: Deploy pre-trained large models to optimize risk identification rules through prompt engineering and retrieval enhancement generation techniques;

[0128] In step S4:

[0129] End-to-end fusion of multimodal e-commerce data using the CommerceMM model;

[0130] Retrieve the historical regulatory case database based on the RAG technology to enhance the generation of risk response strategies;

[0131] Inject domain knowledge into the large model parameters through supervised fine-tuning technology.

[0132] S5: Real-time monitor trading behaviors, and combine the agent strategy and the large model to assist in decision-making to output the risk level and regulatory instructions;

[0133] Step S5 further includes:

[0134] Establish an interpretability analysis module to visually display the risk decision basis through attention weights;

[0135] Support the regulatory personnel to interactively adjust the agent strategy parameters to achieve human-machine collaborative optimization.

[0136] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent assessment system for online transaction credit risk based on multimodal data fusion and deep learning, characterized by: include: Multimodal data collection module, used to collect text, images, videos, audio and user behavior data of online trading platforms in real time; The multimodal data fusion module performs semantic alignment and cross-modal interactive learning on multimodal data through feature-level fusion, model-level fusion, and decision-level fusion strategies to generate a unified multimodal feature vector. The deep reinforcement learning agent module is built on the Actor-Critic network framework, which optimizes the monitoring and supervision strategy through dynamic data interaction with the environment, including the Critic network for evaluating state-action value and the Actor network for generating optimal supervision actions; The large model interaction module integrates pre-trained large models and provides multimodal feature extraction, virtual environment simulation and auxiliary decision support through prompt engineering, retrieval enhancement generation and supervised fine-tuning technology; The risk assessment and supervision execution module identifies violations and generates risk warning signals based on the strategies output by the intelligent agent, and performs differentiated supervision actions.

2. The network transaction credit risk intelligent assessment system based on multimodal data fusion and deep learning according to claim 1 is characterized in that: The multimodal data fusion module uses the following method to achieve data alignment: Capturing the semantic relationship between text and image through cross-modal attention mechanism; Interpolation of missing modality data using masked self-supervised learning; Entity alignment and sentiment polarity matching are achieved based on multimodal embedding technology.

3. The network transaction credit risk intelligent assessment system based on multimodal data fusion and deep learning according to claim 1 is characterized in that: In the deep reinforcement learning agent module: The Critic network uses a deep Q network (DQN) combined with temporal difference learning (TDL) to optimize the value function; The Actor network maximizes long-term cumulative rewards through the policy gradient method and introduces a hierarchical task decomposition mechanism; The experience pool uses priority replay technology to store interaction data, improving the diversity of training samples and the stability of strategies.

4. The network transaction credit risk intelligent assessment system based on multimodal data fusion and deep learning according to claim 1 is characterized in that: The large model interaction module includes: The visual-language pre-trained model based on Kaleido-BERT extracts joint features of product text and images; Use large models to generate simulated transaction data and user behavior sequences to build a reinforcement learning training environment; Through the mind chain technology, real-time monitoring data is combined with historical risk patterns to generate explainable decision recommendations.

5. The network transaction credit risk intelligent assessment system based on multimodal data fusion and deep learning according to claim 1 is characterized in that: The risk assessment and supervision execution module includes: Identify false advertising, fake orders, and sales of prohibited items through multimodal feature matching; Detect big data price discrimination and price violations based on user behavior sequence analysis; Dynamically adjust the regulatory policy threshold to achieve a balanced optimization of the compliance rate and the violation identification recall rate.

6. An evaluation method for an intelligent evaluation system for online transaction credit risk based on multimodal data fusion and deep learning according to any one of claims 1 to 5, characterized in that: The following steps are involved: S1: Collect multimodal data from online trading platforms, clean them, unify the format and align the time series; S2: Generate multimodal joint feature vectors through feature-level fusion and cross-modal attention mechanism; S3: Build a deep reinforcement learning agent and train the Actor-Critic network using simulated environment data and real interaction data; S4: Deploy a large pre-trained model to optimize risk identification rules through prompt engineering and retrieval enhancement generation technology; S5: Monitor trading behaviors in real time, combine intelligent agent strategies with large models to assist decision-making and output risk levels and regulatory instructions.

7. The method for intelligent assessment of credit risk of online transactions based on multimodal data fusion and deep learning according to claim 6 is characterized in that: The training process in step S3 includes: Design a dual reward mechanism to provide positive incentives for compliance and impose penalties on violations; The target network delayed update strategy is used to stabilize the parameter optimization of the critic network; Respond to dynamic changes in trading environment through adaptive learning rate adjustment module.

8. The method for intelligent assessment of credit risk of online transactions based on multimodal data fusion and deep learning according to claim 6 is characterized in that: In step S4: Use the CommerceMM model to perform end-to-end fusion of multimodal e-commerce data; Retrieve historical regulatory case libraries based on RAG technology to enhance the generation of risk response strategies; Infuse domain knowledge into large model parameters through supervised fine-tuning techniques.

9. The method for intelligent assessment of credit risk of online transactions based on multimodal data fusion and deep learning according to claim 6 is characterized in that: The step S5 further comprises: Establish an interpretable analysis module to visualize the basis of risk decision-making through attention weights; It supports supervisors to interactively adjust the strategy parameters of intelligent agents to achieve human-machine collaborative optimization.

Images