A bank flow risk assessment method and application system

Abstract

The application discloses a bank flow risk assessment method and an application system. The method comprises the following steps: acquiring first multi-modal data; wherein the first multi-modal data comprises bank flow, device information corresponding to the bank flow, user portrait corresponding to the device information, social network relationship, and geographic location; inputting the first multi-modal data into a risk score model to obtain a risk level corresponding to the first multi-modal data; and performing risk monitoring according to the risk level. In the foregoing manner, the accuracy and efficiency of a financial institution in identifying the state of enterprise fund flow and potential risks can be improved.

Description

Technical Field

[0001] This application relates to the field of bank transaction risk assessment technology, and in particular to bank transaction risk assessment methods and application systems. Background Technology

[0002] Bank transaction data is an important source of information reflecting the economic activities of businesses and individuals. For financial institutions, including banks, investment banks, securities firms, and auditors, analyzing bank transaction data provides a comprehensive understanding of a company's operations, profitability, and income and expenditure, and can help identify potential misconduct by the target company, thus effectively identifying potential risks.

[0003] Currently, risk assessment during due diligence primarily relies on the following methods:

[0004] (1) Expert scoring method: It relies on expert experience and has problems such as strong subjectivity and poor consistency;

[0005] (2) Credit scoring model: It is based on historical data, but it is difficult to cover all risk scenarios;

[0006] (3) Financial analysis methods: focus on the company’s financial situation, but are difficult to reflect real-time risks.

[0007] Characteristics of bank statement data: Bank statement data has the following characteristics:

[0008] (1) Real-time: Reflects the real-time economic activities of enterprises and individuals;

[0009] (2) Comprehensiveness: Covers multiple accounts and various transaction types;

[0010] (3) Relevance: It is related to other financial data of the enterprise, personal credit status, etc.

[0011] Existing technologies have the following limitations in assessing bank transaction risks:

[0012] (1) Lack of evaluation methods tailored to the characteristics of bank transaction data;

[0013] (2) The diversity of banks makes it difficult to adapt and collect all bank statement formats;

[0014] (3) Banking institutions cannot access the transaction records of other banks through the internet, as all data is provided by the enterprises themselves, making it difficult to verify the integrity and authenticity of the data;

[0015] (4) Unable to effectively identify abnormal transactions and potential risks. Summary of the Invention

[0016] The bank transaction risk assessment method and application system provided in this application can improve the accuracy and efficiency of financial institutions in identifying the status of corporate cash flow and potential risks.

[0017] In a first aspect, this application provides a method for assessing bank transaction risk, the method comprising: acquiring first multimodal data; wherein the first multimodal data includes bank transaction records, device information corresponding to the bank transaction records, user profiles corresponding to the device information, social network relationships, and geographical location; inputting the first multimodal data into a risk scoring model to obtain the risk level corresponding to the first multimodal data; and performing risk monitoring based on the risk level.

[0018] The risk scoring model includes a feature extraction module, a relation learning module, a reinforcement learning module, a generative adversarial network, and a risk scoring engine. Multimodal data is input into the risk scoring model to obtain the risk level corresponding to the multimodal data, including: inputting the first multimodal data into the feature extraction module to obtain a first feature; inputting the first feature into the relation learning module to obtain a second feature; inputting the first feature into the generative adversarial network to obtain a third feature; inputting the first and second features into the reinforcement learning module to obtain a fourth feature; and inputting the first, second, third, and fourth features into the risk scoring engine to obtain the risk level corresponding to the first multimodal data.

[0019] The method further includes: acquiring second multimodal data; the second multimodal data includes data from regulatory agencies, first bank, second bank, and financial institutions; inputting the second multimodal data into an anti-fraud model to obtain target fraud information corresponding to the second multimodal data; inputting the target fraud information into a smart contract risk control module to obtain a risk level; and inputting the target fraud information into a heterogeneous blockchain network for storage; and obtaining a regulatory report based on the risk level.

[0020] The fraud prevention model includes a federated learning privacy protection layer, a fraud prevention layer, and an explainable AI layer. The second multimodal data is input into the fraud prevention model to obtain fraud information corresponding to the second multimodal data. This includes: inputting the second multimodal data into the federated learning privacy protection layer to obtain privacy-protected target multimodal data; inputting the target multimodal data into the fraud prevention layer to obtain predicted fraud information; and inputting the predicted fraud information into the explainable AI layer to obtain target fraud information.

[0021] The method further includes: acquiring third-mode multimodal data; wherein the third-mode multimodal data includes basic user information, social media data, e-commerce behavior data, bank statements, and geographical location; and inputting the third-mode multimodal data into a recommendation model to obtain personalized recommendations corresponding to the third-mode multimodal data.

[0022] The recommendation model comprises: a federated learning privacy protection layer, a user profile construction layer, a deep learning layer, and a personalized recommendation layer. The model inputs third-dimensional multimodal data into the recommendation model to obtain personalized recommendations corresponding to the third-dimensional multimodal data. This process includes: inputting the third-dimensional multimodal data into the federated learning privacy protection layer to obtain privacy-protected target multimodal data; inputting the target multimodal data into the user profile construction layer to obtain user profile features; inputting the user profile features into the deep learning layer to obtain a recommendation strategy; and inputting the recommendation strategy into the personalized recommendation layer to obtain personalized recommendations corresponding to the third-dimensional multimodal data.

[0023] The method further includes: acquiring fourth multimodal data; wherein the fourth multimodal data includes online user behavior, offline transaction data, social media interaction data, customer service logs, and application data; and inputting the fourth multimodal data into a marketing decision model to obtain the marketing decision corresponding to the fourth multimodal data.

[0024] The method further includes: acquiring fifth multimodal data; wherein the fifth multimodal data includes business department requirements, historical resource usage data, risk classification data, external market signals and real-time performance indicators; and inputting the fifth multimodal data into the intelligent resource scheduling model to obtain the resource scheduling decision corresponding to the fifth multimodal data.

[0025] The method further includes: acquiring sixth multimodal data; wherein the sixth multimodal data includes bank statements, news text analysis, social media sentiment, macroeconomic indicators, industry reports and regulatory policy documents; and inputting the sixth multimodal data into a capital flow prediction model to obtain resource scheduling decisions corresponding to the sixth multimodal data.

[0026] Secondly, this application provides a bank transaction risk assessment application system, which includes a processor and a memory connected to the processor; the memory is used to store a computer program, which, when executed by the processor, is used to implement the method provided in the first aspect.

[0027] The beneficial effects of this application are as follows: Unlike existing technologies, the bank transaction risk assessment method and application system provided in this application acquire first multimodal data, which includes bank transactions, corresponding device information, user profiles, social network relationships, and geographical location. This first multimodal data is input into a risk scoring model to obtain the risk level corresponding to the first multimodal data. Risk monitoring is then conducted based on the risk level, overcoming the current data collection problems related to various formats of domestic and international bank transactions. Furthermore, based on the characteristics of bank transaction data and combined with other dimensions of multimodal data, cross-validation is performed to construct a risk assessment model. This model identifies the completeness and authenticity of the transaction records provided by enterprises, and can merge and summarize various transaction relationships to identify fraudulent flows of transactions suspected of being related to them. This improves the accuracy and efficiency of financial institutions in identifying the status of enterprise cash flow and potential risks. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0029] Figure 1 This is a flowchart illustrating an embodiment of the bank transaction risk assessment method provided in this application;

[0030] Figure 2 This is a flowchart illustrating an embodiment of step 12 provided in this application;

[0031] Figure 3 This is a flowchart illustrating another embodiment of the bank statement risk assessment method provided in this application;

[0032] Figure 4 This is a flowchart illustrating an embodiment of step 32 provided in this application;

[0033] Figure 5 This is a flowchart illustrating another embodiment of the bank statement risk assessment method provided in this application;

[0034] Figure 6 This is a flowchart illustrating an embodiment of step 52 provided in this application;

[0035] Figure 7 This is a flowchart illustrating another embodiment of the bank statement risk assessment method provided in this application;

[0036] Figure 8 This is a flowchart illustrating another embodiment of the bank statement risk assessment method provided in this application;

[0037] Figure 9 This is a flowchart illustrating another embodiment of the bank statement risk assessment method provided in this application;

[0038] Figure 10 This is a flowchart illustrating another embodiment of the bank statement risk assessment method provided in this application;

[0039] Figure 11 This is a flowchart illustrating another embodiment of the bank statement risk assessment method provided in this application;

[0040] Figure 12 This is a schematic diagram of an embodiment of the bank transaction risk assessment application system provided in this application. Detailed Implementation

[0041] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only for explaining this application and not for limiting it. Furthermore, it should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all structures. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0042] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0043] Bank transaction data is an important source of information reflecting the economic activities of businesses and individuals. For financial institutions, including banks, investment banks, securities firms, and auditors, analyzing bank transaction data provides a comprehensive understanding of a company's operations, profitability, and income and expenditure, and can help identify potential misconduct by the target company, thus effectively identifying potential risks.

[0044] Currently, in their due diligence and risk assessment, they primarily rely on the following methods:

[0045] (1) Expert scoring method: It relies on expert experience and has problems such as strong subjectivity and poor consistency;

[0046] (2) Credit scoring model: It is based on historical data, but it is difficult to cover all risk scenarios;

[0047] (3) Financial analysis methods: focus on the company’s financial situation, but are difficult to reflect real-time risks.

[0048] Characteristics of bank statement data: Bank statement data has the following characteristics:

[0049] (1) Real-time: Reflects the real-time economic activities of enterprises and individuals;

[0050] (2) Comprehensiveness: Covers multiple accounts and various transaction types;

[0051] (3) Relevance: It is related to other financial data of the enterprise, personal credit status, etc.

[0052] Existing technologies have the following limitations in assessing bank transaction risks:

[0053] (1) Lack of evaluation methods tailored to the characteristics of bank transaction data;

[0054] (2) The diversity of banks makes it difficult to adapt and collect all bank statement formats;

[0055] (3) Banking institutions cannot access the transaction records of other banks through the internet, as all data is provided by the enterprises themselves, making it difficult to verify the integrity and authenticity of the data;

[0056] (4) Unable to effectively identify abnormal transactions and potential risks.

[0057] Based on this, this application proposes to acquire first multimodal data, which includes bank statements, corresponding device information, user profiles, social network relationships, and geographical location. The first multimodal data is then input into a risk scoring model to obtain the risk level corresponding to the first multimodal data. A risk monitoring method is implemented based on the risk level, overcoming the current data collection problems of various formats of domestic and foreign bank statements. Based on the characteristics of bank transaction data, and combined with other dimensions of multimodal data, data cross-validation is performed to construct a risk assessment model. This model identifies the completeness and authenticity of the bank statement records provided by enterprises, and can merge and summarize various transaction relationships to identify fraudulent flows of transactions suspected of being related to them, thereby improving the accuracy and efficiency of financial institutions in identifying the cash flow status and potential risks of enterprises. See any of the following embodiments or any combination of embodiments for details.

[0058] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0059] Step 11: Obtain the first multimodal data; wherein, the first multimodal data includes bank statements, device information corresponding to the bank statements, user profiles corresponding to the device information, social network relationships, and geographical location.

[0060] In some embodiments, a first multimodal data can be obtained by fusing different data sources using a multimodal input layer.

[0061] Step 12: Input the first multimodal data into the risk scoring model to obtain the risk level corresponding to the first multimodal data.

[0062] In one application scenario, the risk scoring model includes: a feature extraction module, a relation learning module, a reinforcement learning module, a generative adversarial network, and a risk scoring engine. (See also...) Figure 2 Step 12 can be specifically described as follows:

[0063] Step 121: Input the first multimodal data into the feature extraction module to obtain the first feature.

[0064] In some embodiments, the feature extraction module includes a ResNet feature extractor and an attention mechanism layer. First multimodal data is input to the corresponding features of the ResNet feature extractor, and then the corresponding features are input to the attention mechanism layer to obtain the first features. The ResNet feature extractor is capable of capturing complex nonlinear features.

[0065] Step 122: Input the first feature into the relation learning module to obtain the second feature.

[0066] In some embodiments, the relationship learning module includes a graph neural network and a relationship embedding layer. A first feature is input into the graph neural network, which constructs a graph of transaction relationships from the first multimodal data. This graph is then input into the relationship embedding layer to obtain a second feature. The graph neural network is capable of learning relationships between accounts.

[0067] Step 123: Input the first feature into the generative adversarial network to obtain the third feature.

[0068] In some embodiments, a generative adversarial network (GAN) includes a generator and a discriminator. A first feature is input to the generator to obtain a corresponding feature, which is then input to the discriminator. GANs can enhance the generalization ability of a model.

[0069] Step 124: Input the first and second features into the reinforcement learning module to obtain the fourth feature.

[0070] In some embodiments, the reinforcement learning module includes a policy network, a value network, and a reward function. A first feature and a second feature are input into the policy network to obtain a corresponding policy. This policy is then input into the value network to evaluate the value of each policy. Finally, the policy and the value are input into the reward function to calculate the reward, thereby obtaining a fourth feature. The reinforcement learning module is used to dynamically optimize the detection policy.

[0071] Step 125: Input the first feature, second feature, third feature and fourth feature into the risk scoring engine to obtain the risk level corresponding to the first multimodal data.

[0072] In some embodiments, the risk scoring engine includes a multilayer perceptron, a confidence scoring module, and a risk classification module. First, second, third, and fourth features are input into the multilayer perceptron to obtain the perceptual features output by the multilayer perceptron. The confidence scoring module calculates the confidence level of the perceptual features to obtain the corresponding confidence score. The risk classification module then classifies the perceptual features to obtain the corresponding risk type. Based on this risk type and the corresponding confidence score, the risk level corresponding to the first multimodal data is obtained. The multilayer perceptron can comprehensively assess risk.

[0073] Step 13: Conduct risk monitoring based on the risk level.

[0074] In some embodiments, risk levels can be categorized as low risk, medium risk, and high risk. At low risk, the application is allowed to proceed normally. At medium risk, manual review is required. At high risk, the account is frozen and / or an alert is issued.

[0075] Furthermore, the risk scoring model is continuously learned and optimized based on the risk level.

[0076] In this embodiment, first multimodal data is acquired, including bank statements, corresponding device information, user profiles, social network relationships, and geographic location. This first multimodal data is input into a risk scoring model to obtain the risk level corresponding to the first multimodal data. Risk monitoring is then implemented based on this risk level, overcoming the current data collection problems related to various formats of domestic and international bank statements. Based on the characteristics of bank transaction data, and combined with other dimensions of multimodal data, cross-validation is performed to construct a risk assessment model. This model identifies the completeness and authenticity of the bank statement records provided by the enterprise, and can merge and summarize various transaction relationships to identify fraudulent flows of transactions suspected of being related to the enterprise. This improves the accuracy and efficiency of financial institutions in identifying the status of enterprise cash flow and potential risks. This embodiment can be applied to deep learning-based abnormal behavior detection systems.

[0077] See Figure 3 , Figure 3 This is a flowchart illustrating another embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0078] Step 31: Obtain the second multimodal data; the second multimodal data includes data from regulatory agencies, first bank data, second bank data, and financial institutions.

[0079] Step 32: Input the second multimodal data into the anti-fraud model to obtain the target fraud information corresponding to the second multimodal data.

[0080] In one application scenario, the fraud prevention model includes a federated learning privacy protection layer, a fraud prevention layer, and an explainable AI layer. (See also...) Figure 4 Step 32 can be the following process:

[0081] Step 321: Input the second multimodal data into the federated learning privacy protection layer to obtain the privacy-protected target multimodal data.

[0082] In some embodiments, the federated learning privacy protection layer includes a federated learning aggregator, a homomorphic encryption module, and a differential privacy mechanism. The second multimodal data is input into the federated learning aggregator to obtain federated learning features, which are then encrypted using the homomorphic encryption module to obtain encrypted features. The differential privacy mechanism is then used to process the encrypted features to obtain the privacy-protected target multimodal data.

[0083] Step 322: Input the target multimodal data into the fraud prevention layer to obtain the predicted fraud information.

[0084] In some embodiments, the anti-fraud layer includes a feature extraction module, a graph neural network, a generative adversarial network (GAN), and a reinforcement learning policy network. Target multimodal data is input to the feature extraction module to obtain multimodal features. These multimodal features are then input to the graph neural network to obtain relational graph features. The relational graph features are further enhanced by the GAN, and the enhanced relational graph features are then input to the reinforcement learning policy network for policy optimization to obtain predicted fraud information. The feature extraction module can be composed of a ResNet feature extractor. That is, the anti-fraud layer can perform operations such as ResNet feature extraction, graph neural network relation learning, GAN enhancement, and reinforcement learning dynamic optimization.

[0085] Step 323: Input the predicted fraud information into the interpretable AI layer to obtain the target fraud information.

[0086] In some embodiments, the interpretable AI layer includes a decision tree visualization module, a feature contribution decomposer, and a local interpretability builder. Predicted fraud information is input to the decision tree visualization module to visualize the decision-making process. The predicted fraud information is also input to the feature contribution decomposer to obtain first target fraud information. Finally, the predicted fraud information is input to the local interpretability builder to obtain second target fraud information.

[0087] Step 33: Input the target fraud information into the smart contract risk control module to obtain the risk level; and input the target fraud information into the heterogeneous blockchain network for storage.

[0088] In some embodiments, the smart contract risk control module includes a risk assessment contract module, an automatic early warning mechanism, and a transaction freezing module. Target fraud information is input into the risk assessment contract module to obtain risk information. The automatic early warning mechanism then automatically issues an early warning based on the risk information. The transaction freezing module then freezes transactions based on the early warning results and performs a final risk score to determine the risk level. Abnormal transactions are also marked.

[0089] Heterogeneous blockchain networks include modules such as consortium blockchains, public connection interfaces, private blockchain interactions, and cross-chain protocols, which can interconnect different blockchain networks to build a more secure distributed ledger system.

[0090] Step 34: Obtain a regulatory report based on the risk level.

[0091] This embodiment introduces federated learning technology to achieve multi-party collaborative modeling and improve the model's generalization ability while protecting data privacy. It also combines interpretable AI technology to explain the model's predictions, enhancing its transparency and credibility. Furthermore, it designs smart contracts to automate risk assessment and response, reducing human intervention. Overall, it achieves a comprehensive integration of data security, model intelligence, and accurate risk identification, providing a new technological paradigm for financial anti-fraud. This embodiment can be applied to anti-fraud systems combining blockchain and AI.

[0092] See Figure 5 , Figure 5 This is a flowchart illustrating another embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0093] Step 51: Obtain third-mode multimodal data; whereby the third-mode multimodal data includes basic user information, social media data, e-commerce behavior data, bank statements, and geographical location.

[0094] Step 52: Input the third multimodal data into the recommendation model to obtain personalized recommendations corresponding to the third multimodal data.

[0095] In one application scenario, this recommendation model includes: a federated learning privacy protection layer, a user profile construction layer, a deep learning layer, and a personalized recommendation layer. (See also...) Figure 6 Step 52 can be the following process:

[0096] Step 521: Input the third multimodal data into the federated learning privacy protection layer to obtain the privacy-protected target multimodal data.

[0097] In some embodiments, the federated learning privacy protection layer includes a federated learning aggregator, a homomorphic encryption module, and a differential privacy mechanism. Third-mode multimodal data is input to the federated learning aggregator to obtain federated learning features, which are then encrypted using the homomorphic encryption module to obtain encrypted features. The differential privacy mechanism is then used to process the encrypted features to obtain the privacy-protected target multimodal data.

[0098] Step 522: Input the target multimodal data into the user profile construction layer to obtain user profile features.

[0099] In some embodiments, the user profile construction layer includes a demographic feature module, an interest tag extraction module, a behavioral preference analysis module, and a user profile embedding module. Target multimodal data is input to the demographic feature module to obtain demographic features. The demographic features are input to the interest tag extraction module to obtain interest tags. The interest tags and demographic features are input to the behavioral preference analysis module to obtain behavioral preference features. The behavioral preference features are input to the user profile embedding module to obtain user profile features.

[0100] Step 523: Input the user profile features into the deep learning layer to obtain the recommendation strategy.

[0101] In some embodiments, the deep learning layer includes: a feature extraction module, a graph neural network, a generative adversarial network (GAN), and a reinforcement learning policy network. User profile features are input to the feature extraction module to obtain multimodal features, which are then input to the graph neural network to obtain relationship graph features. These relationship graph features are then input to the GAN for enhancement, and the enhanced relationship graph features are input to the reinforcement learning policy network for policy optimization to obtain a recommendation policy. The feature extraction module can be composed of a ResNet feature extractor. That is, ResNet feature extraction, graph neural network relationship learning, GAN enhancement, and reinforcement learning dynamic optimization can be performed within the deep learning layer.

[0102] Step 524: Input the recommendation strategy into the personalized recommendation layer to obtain personalized recommendations corresponding to the third multimodal data.

[0103] In some embodiments, the personalized recommendation layer includes a collaborative filtering module, a context-sensitive recommendation module, and a real-time recommendation adjustment module. The recommendation strategy is input to the collaborative filtering module to obtain a filtered recommendation strategy. The filtered recommendation strategy is then input to the context-sensitive recommendation module and the real-time recommendation adjustment module to optimize and adjust the filtered recommendation strategy, resulting in personalized recommendations corresponding to the third multimodal data.

[0104] Furthermore, the personalized recommendations output by the recommendation model can be used to predict click-through rates and score recommendation satisfaction, allowing for continuous optimization and learning of the recommendation model based on the satisfaction scores. Additionally, user interaction feedback can be obtained after outputting personalized recommendations to further optimize the recommendation model.

[0105] In this embodiment, the intelligent recommendation engine utilizes graph neural networks to model the complex relationships between customers and products / services, providing more accurate recommendations. Furthermore, it leverages reinforcement learning to continuously optimize recommendation strategies through user interaction, achieving long-term optimization of personalized recommendations.

[0106] In addition, it incorporates customer profile data such as age, gender, and occupation to improve the accuracy of recommendations. It also integrates external data, including social media and e-commerce data, to enrich user profiles.

[0107] Federated learning enables collaborative modeling of multi-party data while protecting data privacy. Differential privacy is used to add noise to protect user privacy.

[0108] See Figure 7 , Figure 7 This is a flowchart illustrating another embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0109] Step 71: Obtain the fourth multimodal data; wherein, the fourth multimodal data includes online user behavior, offline transaction data, social media interaction data, customer service logs, and application data.

[0110] After acquiring the fourth multimodal data, it can be processed using streaming computing and event processing modules. For example, a scalable distributed streaming platform, a distributed streaming processing system, a real-time event analysis engine, and an event-triggered rule engine can be used to process the fourth multimodal data.

[0111] Step 72: Input the fourth multimodal data into the marketing decision model to obtain the marketing decision corresponding to the fourth multimodal data.

[0112] In some embodiments, the marketing decision model includes a user segmentation model, a personalized recommendation module, a marketing strategy generator, and a natural language generation module. The fourth multimodal data is input into the user segmentation model to obtain corresponding user segmentation results. These user segmentation results are then input into the personalized recommendation module and the marketing strategy generator to obtain corresponding strategy features. Finally, the strategy features are input into the natural language generation module to obtain the marketing decision corresponding to the fourth multimodal data.

[0113] Furthermore, implement marketing decisions through multi-channel marketing strategies, such as online advertising, SMS / email marketing, personalized pages / content, and intelligent customer service recommendations.

[0114] Furthermore, the marketing effectiveness should be evaluated. This can be done using techniques such as causal inference models, data-driven decision-making frameworks, A / B testing, marketing ROI analysis, and in-depth conversion rate analysis.

[0115] Furthermore, continuous optimization mechanisms are employed to improve the overall system, such as reinforcement learning optimizers, automatic model tuning, and marketing strategy evolution. This leads to feedback on the marketing strategy.

[0116] In this embodiment, streaming computing enables real-time processing of massive amounts of data and timely responses to customer behavior. An event-driven architecture allows for event-triggered marketing campaigns, improving response speed. Multi-channel marketing breaks down bank transactions into online and offline records, integrating online and offline channels to achieve omnichannel marketing and enhance marketing effectiveness. Personalized content generation, based on natural language processing technology, generates customized content for different channels. Marketing effectiveness evaluation utilizes causal inference to assess the true impact of marketing campaigns on user behavior. A / B testing is used to compare the effectiveness of different marketing strategies and optimize them.

[0117] In other words, by breaking down data barriers and introducing intelligent technologies, the system has achieved a new data-driven marketing paradigm for the future, providing more accurate, personalized, and efficient marketing solutions. This embodiment can be applied to data-driven marketing automation systems.

[0118] See Figure 8 , Figure 8 This is a flowchart illustrating another embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0119] Step 81: Obtain the fifth multimodal data; the fifth multimodal data includes business department requirements, historical resource usage data, risk classification data, external market signals, and real-time performance indicators.

[0120] Step 82: Input the fifth multimodal data into the intelligent resource scheduling model to obtain the resource scheduling decision corresponding to the fifth multimodal data.

[0121] In some embodiments, the intelligent resource scheduling model includes a deep learning intelligent analysis layer, a multi-objective optimization engine, a risk control module, and an intelligent scheduling decision layer.

[0122] The fifth multimodal data is input into the deep learning intelligent analysis layer to generate a preliminary scheduling strategy. This preliminary scheduling strategy is then input into the multi-objective optimization engine to obtain the corresponding resource allocation weights. The preliminary scheduling strategy and resource allocation weights are input into the risk control module to obtain the corresponding intermediate scheduling strategy. Finally, the resource allocation weights and intermediate scheduling strategy are input into the intelligent scheduling decision layer to obtain the resource scheduling decision corresponding to the fifth multimodal data.

[0123] The deep learning intelligent analysis layer includes a sequence prediction module, a feature extraction module, a graph neural network, a reinforcement learning policy network, and a generative adversarial network (GAN). The fifth multimodal data is input into the sequence prediction module to obtain the target sequence. The target sequence is then input into the feature extraction module to obtain multimodal features. These multimodal features are then input into the graph neural network to obtain relational graph features, which are then input into the reinforcement learning policy network for policy optimization. The optimized policy is then input into the GAN for enhancement and initial policy scheduling. The feature extraction module can be constructed using a ResNet feature extractor.

[0124] The multi-objective optimization engine includes a Pareto optimality solution module, a multi-objective genetic algorithm module, a constraint evaluation module, and a resource allocation weight calculation module. Specifically, these modules process the initial scheduling strategy to obtain the corresponding resource allocation weights.

[0125] The risk control module includes an anomaly detection algorithm unit, a risk assessment model, a risk early warning mechanism, and an emergency resource allocation unit. Specifically, by utilizing the anomaly detection algorithm unit, risk assessment model, risk early warning mechanism, and emergency resource allocation unit, combined with the preliminary scheduling strategy and resource allocation weights, a corresponding intermediate scheduling strategy is obtained.

[0126] The intelligent scheduling decision layer includes a resource allocation strategy module, a real-time scheduling execution module, a dynamic load balancing module, and a resource utilization optimization module. Specifically, these modules process the input resource allocation weights and intermediate scheduling strategies to obtain the resource scheduling decision corresponding to the fifth multimodal data.

[0127] Furthermore, the model can be continuously learned and optimized by utilizing performance feedback loop mechanisms, automatic model tuning mechanisms, and decision knowledge accumulation mechanisms, thereby obtaining resource scheduling optimization reports for decision-makers to refer to and make decisions.

[0128] In this embodiment, the above method can be used in an intelligent resource scheduling system. The sequence prediction module can be composed of a sequence model (RNN / LSTM), capable of considering time-series data, predicting future resource demands, and achieving more accurate scheduling. Reinforcement learning is used to transform the resource scheduling problem into a Markov decision process, finding the optimal scheduling strategy through trial and error. Graph neural networks are used to model the complex relationships between various business departments within the bank, achieving finer-grained resource allocation. Pareto optimization is used to simultaneously consider multiple optimization objectives, such as maximizing revenue, minimizing risk, and balancing load. Multi-objective genetic algorithms are used to find a set of non-dominated solutions, providing decision-makers with more choices. Anomaly detection algorithms are used to identify abnormal resource usage and take timely measures. A risk assessment model is used to evaluate the risks under different scheduling strategies, ensuring stable system operation. In short, the overall intelligent resource scheduling system can optimize bank resource allocation, improve capital utilization efficiency, and reduce operating costs.

[0129] See Figure 9 , Figure 9 This is a flowchart illustrating another embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0130] Step 91: Obtain the sixth multimodal data; the sixth multimodal data includes bank statements, news text analysis, social media sentiment, macroeconomic indicators, industry reports, and regulatory policy documents.

[0131] In some embodiments, the sixth multimodal data can be cleaned using multimodal data cleaning, characterized using feature engineering, and then standardized and embedded to obtain the final features. These final features are then input into the capital flow prediction model.

[0132] Step 92: Input the sixth multimodal data into the capital flow prediction model to obtain the resource scheduling decision corresponding to the sixth multimodal data.

[0133] The cash flow prediction model includes a deep learning prediction module, a causal inference and analysis module, a dynamic model update module, a prediction and risk assessment module, and a visualization and reporting module.

[0134] In some embodiments, the deep learning prediction module includes: a feature extraction module, a graph neural network, a generative adversarial network (GAN), a reinforcement learning prediction policy network, and a long short-term memory (LSTM) network. The sixth multimodal data is input to the feature extraction module to obtain multimodal features. These multimodal features are then input to the graph neural network to obtain relational graph features. These relational graph features are further input to the GAN for enhancement. The enhanced relational graph features are then input to the reinforcement learning prediction policy network for policy optimization to obtain a prediction policy. Finally, the prediction policy is input to the LSM network for temporal modeling to obtain the modeled prediction policy. The feature extraction module can be composed of a ResNet feature extractor. That is, the deep learning prediction module can perform ResNet feature extraction, graph neural network relation learning, GAN enhancement, and reinforcement learning dynamic optimization.

[0135] The causal inference and analysis module includes causal graph construction, counterfactual inference, key influencing factor identification, and scenario simulation engine. The modeled prediction strategy is input into the causal inference and analysis module, which then processes the strategy using these modules to obtain the analysis results.

[0136] The dynamic model update module includes online learning, transfer learning, model performance evaluation, and adaptive parameter adjustment mechanisms. The modeled prediction strategy and analysis results are input into the dynamic model update module, which then processes these mechanisms to update the model and obtain the desired target results.

[0137] The prediction and risk assessment module includes functions such as capital flow prediction, risk scoring, anomaly detection, and decision support. The target results are input into the prediction and risk assessment module to obtain the resource scheduling decision corresponding to the sixth multimodal data.

[0138] Furthermore, after obtaining the resource scheduling decision corresponding to the sixth multimodal data, the visualization and reporting module is used to interpret the prediction results, provide risk warnings and decision suggestions, and visualize the results through an interactive dashboard.

[0139] In this embodiment, a big data-based capital flow prediction system can be applied. Text data can be used to analyze news, reports, and other textual information to capture market trends. Social media data can be used to analyze public opinion on social media to predict market fluctuations. External economic data can be combined with macroeconomic data to improve prediction accuracy.

[0140] Furthermore, it utilizes causal graph models to establish causal relationship diagrams between fund flows and influencing factors, delving into causal mechanisms. It also employs counterfactual reasoning to simulate fund flows under different scenarios, assessing the potential impact of decisions. Additionally, it dynamically updates the model through online learning, ensuring timely predictions as new data continuously enters. Finally, it leverages transfer learning to integrate knowledge from other fields or banking into the current model, accelerating model training.

[0141] See Figure 10 , Figure 10 This is a flowchart illustrating another embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0142] Step 101: Obtain the seventh multimodal data; the seventh multimodal data includes external credit data, bank statements, social media sentiment, news information, market data, and regulatory reports.

[0143] In some embodiments, the seventh multimodal data can be cleaned using multimodal data cleaning, characterized using feature engineering, and then standardized before data embedding is performed to obtain the final features. These final features are then input into the risk analysis model.

[0144] Step 102: Input the seventh multimodal data into the risk analysis model to obtain the risk analysis results corresponding to the seventh multimodal data.

[0145] In some embodiments, the risk analysis model includes a deep learning prediction module, a risk detection and assessment module, a real-time early warning mechanism module, a visualization and interaction module, and a continuous learning and optimization module.

[0146] In some embodiments, the deep learning prediction module includes: a feature extraction module, a graph neural network, a generative adversarial network (GAN), a reinforcement learning risk policy network, and a multi-dimensional risk assessment model. The seventh multimodal data is input to the feature extraction module to obtain multimodal features. These multimodal features are then input to the graph neural network to obtain relational graph features. These relational graph features are further input to the GAN for enhancement. The enhanced relational graph features are then input to the reinforcement learning risk policy network for policy optimization to obtain a prediction policy. The prediction policy is then input to the multi-dimensional risk assessment model to obtain preliminary risk results. The feature extraction module can be composed of a ResNet feature extractor. That is, the deep learning prediction module can perform operations such as ResNet feature extraction, graph neural network relation learning, GAN enhancement, and reinforcement learning dynamic optimization.

[0147] In some embodiments, the risk detection and assessment module includes a credit risk model, a market risk model, an operational risk model, a reputational risk model, and an anomaly detection model. Preliminary risk results are input into the risk detection and assessment module, which then analyzes these results using the credit risk model, market risk model, operational risk model, reputational risk model, and anomaly detection model to obtain assessment results for the corresponding risk type.

[0148] The real-time early warning mechanism module includes a risk scoring engine, an alarm triggering rule module, a priority classification module, and an automatic response process module. The risk type assessment results are input into the real-time early warning mechanism module. The risk scoring engine, alarm triggering rule module, priority classification module, and automatic response process module process the risk type assessment results to obtain the risk analysis results corresponding to the seventh multimodal data.

[0149] The visualization and interaction module includes functions such as interactive dashboards, geographic information visualization, risk heat maps, multi-dimensional charts, and risk trend analysis. After obtaining the risk analysis results corresponding to the seventh multimodal data, the interactive dashboards, geographic information visualization, risk heat maps, multi-dimensional charts, and risk trend analysis functions are used to visualize the risk analysis results.

[0150] The continuous learning and optimization module includes functions such as model performance evaluation, automatic parameter tuning, knowledge graph updating, and decision feedback loop, which are used for continuous learning and optimization.

[0151] In this embodiment, it can be applied to a real-time data monitoring platform. Within this platform, the data source can be expanded beyond traditional bank transaction data to include external data sources such as social media, news reports, and market trends, thus building a more comprehensive risk monitoring system.

[0152] In addition, risk model optimization is carried out by using multi-dimensional risk models to construct risk models that cover credit risk, market risk, operational risk, reputational risk and other dimensions.

[0153] Furthermore, it utilizes dynamic risk assessment methods, adjusting risk model parameters dynamically based on real-time data to improve model adaptability. It also employs anomaly detection algorithms to promptly identify abnormal fluctuations in the data, such as large transactions or frequent operations.

[0154] It also enhances visualization by providing a user-friendly interactive dashboard that supports customizable views and alarm settings. Furthermore, it visualizes geographic information, displaying risk events on a map to intuitively show the distribution of risks.

[0155] See Figure 11 , Figure 11This is a flowchart illustrating another embodiment of the bank transaction risk assessment method provided in this application. The method includes:

[0156] Step 111: Obtain the eighth multimodal data; wherein, the eighth multimodal data includes social network data, user images, text and document data, and bank statements.

[0157] Step 112: Input the social network data into the deep learning model to obtain the first feature.

[0158] In some embodiments, the GNN module in a deep learning model can be used to learn relationships in social network data to obtain first features. Relationships can then be mined from social network data to understand complex interest relationships or behavioral patterns.

[0159] Step 113: Input the user's image into the deep learning model to obtain the second feature.

[0160] In some embodiments, the ResNet feature extraction module in a deep learning model can be used to extract features from user images to obtain second features. These features can be extracted to provide visual support for compliance analysis (e.g., authentication or fraud detection).

[0161] Step 114: Input the text and document data, bank statements, first feature and second feature into the compliance analysis module to obtain the compliance scenario.

[0162] In some embodiments, text and document data can be processed using natural language and then input into the compliance analysis module.

[0163] In some embodiments, text and document data, bank statements, and first and second features are input into a compliance analysis module to construct a compliance knowledge graph. This graph is then enhanced with natural language processing, including multilingual support, causal relationship analysis, and sentiment analysis. The enhanced data is then subjected to text and knowledge compliance analysis to enrich the compliance context. Finally, scenario simulation is performed to obtain compliance scenarios. Reinforcement learning is used to optimize compliance strategies for business processes. Furthermore, GANs are used to generate high-quality virtual data and scenarios to supplement insufficient real-world data.

[0164] Step 115: Input the compliance scenario into the risk scoring engine to obtain a compliance risk assessment.

[0165] In some embodiments, compliance scenarios can be input into the reinforcement learning module of a deep learning model for reinforcement learning, and the data after reinforcement learning can be input into a risk scoring engine for real-time compliance risk assessment to obtain a compliance risk assessment. The reinforcement learning module optimizes decisions in simulated scenarios and predicts potential risks. Combining this with virtual data generated by GANs improves the accuracy of model evaluation.

[0166] In this embodiment, the above method can be applied to an intelligent compliance management system. Specifically, by utilizing a compliance knowledge graph, a comprehensive compliance knowledge graph covering laws, regulations, regulatory policies, and industry standards can be established, improving the accuracy of compliance analysis. Multi-language support enables the analysis of compliance documents in multiple languages. Sentiment analysis is used to analyze the sentiment trends in text to identify potential compliance risks. Causal relationship analysis is used to uncover causal relationships in text, gaining a deeper understanding of the causes of compliance events. Furthermore, compliance scenario simulation is used to construct virtual compliance scenarios, conducting simulated tests on different business processes to proactively identify potential compliance issues.

[0167] Furthermore, to improve the speed and accuracy of banking data processing and achieve intelligent analysis and automated processing of transaction data, the following methods can be adopted:

[0168] First, in the multimodal data section, in addition to traditional bank transaction data, we will consider integrating external data sources (such as social media, third-party payment, etc.) to enrich the data dimensions and enhance the depth of analysis.

[0169] Secondly, it can form a distributed data processing system based on cloud computing.

[0170] Introducing real-time stream processing technology into distributed data processing systems enables rapid processing of massive amounts of real-time data, achieving real-time monitoring and early warning. Furthermore, leveraging edge computing decentralizes some computational tasks to edge nodes, reducing network latency and improving system response speed.

[0171] At the same time, pay attention to data security and privacy protection: adopt more advanced encryption algorithms and access control mechanisms to ensure data security and comply with relevant regulations.

[0172] Specific technical points may include:

[0173] Distributed storage: HDFS, S3, Ceph, etc. Using distributed storage systems such as HDFS, S3, and Ceph provides efficient and reliable data storage capabilities.

[0174] Distributed computing: Spark, Flink, Hadoop MapReduce, etc. Batch processing tasks use Spark, Flink, or Hadoop MapReduce to perform computations on large-scale offline data.

[0175] Stream processing: Kafka, Storm, etc. Real-time data stream processing uses Kafka or Storm to achieve fast response and real-time monitoring.

[0176] Data security includes encryption algorithms (AES, RSA, etc.), access control (RBAC), and data masking. Data encryption uses advanced encryption algorithms such as AES and RSA to ensure secure data storage and transmission. Access control can employ role-based access control (RBAC) to restrict access to sensitive data. Data masking can mask sensitive information during data processing to ensure privacy compliance.

[0177] Finally, the analysis results are stored in a database to support business applications and generate compliance reports. The real-time alert module provides immediate risk warnings.

[0178] Specifically, it can be as follows:

[0179] The system acquires heterogeneous data sources, preprocesses them, and inputs the data into the edge computing layer. Edge nodes in the edge computing layer then perform distributed computing and real-time data processing, utilizing a stream processing engine for access control. Real-time monitoring and analysis are integrated with a real-time processing system, and event detection provides real-time alerts, which are then output to business application systems. A data security layer is also used to output compliance reports to business application systems. For distributed computing, the analysis results are stored in a database and fed back to the business application systems.

[0180] In this process, third-party payment data, social media data, and traditional banking data from heterogeneous data sources are distributed and stored, and then distributed computation is performed on this data before it is fed into the stream processing engine. Encryption is used during storage.

[0181] In some embodiments, the distributed computing architecture can adopt a distributed data center approach. The asynchronous federated learning training workflow within the data center is as follows:

[0182] 1. Model Distribution: The container manager selects an idle container for model training. The selection is based on the version of the target model (i.e., the number of training iterations) and the container's historical training logs.

[0183] 2. Resource allocation: The resource manager allocates the most cost-effective hardware resources to the container based on the model version and the container's historical runtime.

[0184] 3. Model Training: The container uses its data to train the received planetary model.

[0185] 4. Model Collection: The container sends the trained planetary model to the model manager, and the container manager updates the container information.

[0186] 5. Model Update: The model manager aggregates the trained planetary and stellar models using specific weights and updates the corresponding planetary models in the model list.

[0187] This asynchronous training process will continue until the training time reaches the threshold.

[0188] The model rotation process between data centers is as follows:

[0189] 1. Model Distribution: When the training time in a data center reaches a threshold, the model manager in each data center aggregates all planetary models to generate a master model, and sends this master model, along with the average number of training iterations of all planetary models, to the target data center. Each data center will take turns selecting a target data center in different rotation rounds.

[0190] 2. Model Routing: The output port transmits the model to the target data center according to the current routing table. The routing table selects the shortest path for model transmission.

[0191] 3. Model Update: After the input port receives the main model, the model manager uses this model to replace the old star model.

[0192] Before the model distribution step, the model manager resets the timer and waits for the next round of model rotation.

[0193] Key step: The stellar model and the master model communicate directly via a model communicator.

[0194] The resource manager is responsible for allocating resources to the training container.

[0195] The container selector selects a model from the planetary models and trains it in the container using the allocated resources.

[0196] This structure ensures asynchronous and efficient model training while reducing communication overhead and the complexity of resource scheduling.

[0197] For example, a distributed computing framework includes Data Center 1, Data Center 2, Data Center 3, and Data Center 4. Each data center involves the updating and distribution of the Stellar model, as well as model rotation, through model distribution, model routing, and model updates, replacing the old Stellar model with a new one.

[0198] In some embodiments, multimodal data may also be acquired. Multimodal data includes business process data, user images, document / text data, and social network data.

[0199] Business process data is input into the unsupervised learning module, which outputs the first feature. User images are input into the ResNet feature extraction module, which outputs the second feature. Document / text data is input into the Transformer model, which outputs the third feature. Social network data is input into the GNN relationship learning module, which outputs the fourth feature.

[0200] The first, second, third, and fourth features are input into the feature fusion module to obtain fused features. Data pattern analysis is performed on these fused features to uncover data patterns and anomalies. Anomaly detection is also performed on the fused features to identify potential data problems. The data pattern analysis results and anomaly detections are then input into the reinforcement learning module to optimize the analysis strategy and generate virtual data.

[0201] The data output by the reinforcement learning module is then used for risk prediction and decision-making, and the interpretable AI module is used to generate analysis reports and provide decision-making basis.

[0202] Furthermore, automated data analysis systems are used to display analysis reports and monitor and warn of risks in real time.

[0203] The above can be applied to automated data analysis systems. Deep learning models incorporate Transformers to enhance their expressive and generalization abilities. Unsupervised learning is used to explore unsupervised learning algorithms and uncover potential patterns and anomalies in the data. Interpretable AI is leveraged to enhance the interpretability of the model, increasing user trust in the model's results.

[0204] Among them, ResNet feature extraction extracts high-level visual features from images for analyzing user behavior or risk.

[0205] Transformer models can handle complex text, improve language understanding, and support multilingual and contextual modeling.

[0206] GNN relation learning analyzes social networks and graph structure data to uncover node relationships and graph characteristics.

[0207] The unsupervised learning module discovers potential patterns and anomalies in the data through clustering and dimensionality reduction.

[0208] The feature fusion module integrates features from different data sources to enhance the model's expressive and generalization capabilities.

[0209] The GAN generation module generates virtual data for simulation testing and data augmentation.

[0210] The reinforcement learning module optimizes strategies and provides dynamic risk prediction and decision-making capabilities.

[0211] The interpretability AI module enhances the interpretability of the model, increasing user trust by generating highly readable analysis reports.

[0212] Automated data analytics systems provide real-time monitoring, data visualization, and alerting capabilities to support automated risk management for businesses.

[0213] See Figure 12 , Figure 12This is a schematic diagram of the structure of an embodiment of the bank transaction risk assessment application system provided in this application. The bank transaction risk assessment application system 120 includes a processor 121 and a memory 122 connected to the processor 121; the memory 122 is used to store a computer program, which, when executed by the processor 121, is used to implement the method of any of the above embodiments.

[0214] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0215] If the integrated units in the other embodiments described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0216] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for assessing bank transaction risk, characterized in that, The method includes: acquiring first multimodal data; wherein the first multimodal data includes bank statements, device information corresponding to the bank statements, user profiles corresponding to the device information, social network relationships, and geographical location; inputting the first multimodal data into a risk scoring model to obtain a risk level corresponding to the first multimodal data; and performing risk monitoring based on the risk level; the risk scoring model includes: a feature extraction module, a relationship learning module, a reinforcement learning module, a generative adversarial network, and a risk scoring engine; the step of inputting the first multimodal data into the risk scoring model to obtain the risk level corresponding to the first multimodal data includes: inputting the first multimodal data into a ResNet feature extractor in the feature extraction module to extract corresponding features, and then inputting the corresponding features into an attention mechanism layer to obtain a first feature; inputting the first feature into a graph neural network in the relationship learning module, and using the graph neural network to analyze the transaction relationships in the first multimodal data. The graph is constructed and input into the relation embedding layer to obtain the second feature. The first feature is input into the generator of the generative adversarial network to obtain the corresponding feature, and then input into the discriminator to obtain the third feature. The first and second features are input into the policy network of the reinforcement learning module to obtain the corresponding policy, and the policy is input into the value network to evaluate the value of each policy. Then, the policy and value are input into the reward function to calculate the reward, resulting in the fourth feature. The first, second, third, and fourth features are input into the multilayer perceptron of the risk scoring engine to obtain the perceptual features output by the multilayer perceptron. The confidence scoring module in the risk scoring engine is used to calculate the confidence of the perceptual features to obtain the corresponding confidence. The risk classification module in the risk scoring engine is used to classify the perceptual features to obtain the corresponding risk type. Based on the risk type and the corresponding confidence, the risk level corresponding to the first multimodal data is obtained.

2. The method according to claim 1, characterized in that, The method further includes: Acquire second multimodal data; the second multimodal data includes regulatory data, bank statements of the first bank, bank statements of the second bank, and data from financial institutions; The second multimodal data is input into the anti-fraud model to obtain the target fraud information corresponding to the second multimodal data; The target fraud information is input into the smart contract risk control module to obtain the risk level; And the target fraud information is input into a heterogeneous blockchain network for storage; A regulatory report will be obtained based on the stated risk level.

3. The method according to claim 2, characterized in that, The fraud prevention model includes a federated learning privacy protection layer, a fraud prevention layer, and an explainable AI layer; the step of inputting the second multimodal data into the fraud prevention model to obtain the fraud information corresponding to the second multimodal data includes: The second multimodal data is input into the federated learning privacy protection layer to obtain the privacy-protected target multimodal data; The target multimodal data is input into the anti-fraud layer to obtain predicted fraud information; The predicted fraud information is input into the explainable AI layer to obtain the target fraud information.

4. The method according to claim 1, characterized in that, The method further includes: Acquire third-mode multimodal data; wherein, the third-mode multimodal data includes basic user information, social media data, e-commerce behavior data, bank statements, and geographical location; The third multimodal data is input into the recommendation model to obtain personalized recommendations corresponding to the third multimodal data.

5. The method according to claim 4, characterized in that, The recommendation model includes: a federated learning privacy protection layer, a user profile construction layer, a deep learning layer, and a personalized recommendation layer; the step of inputting the third multimodal data into the recommendation model to obtain personalized recommendations corresponding to the third multimodal data includes: The third multimodal data is input into the federated learning privacy protection layer to obtain the privacy-protected target multimodal data; The target multimodal data is input into the user profile construction layer to obtain user profile features; The user profile features are input into the deep learning layer to obtain the recommendation strategy; The recommendation strategy is input into the personalized recommendation layer to obtain personalized recommendations corresponding to the third multimodal data.

6. The method according to claim 1, characterized in that, The method further includes: Acquire fourth multimodal data; wherein, the fourth multimodal data includes online user behavior, offline transaction data, social media interaction data, customer service logs, and application data; The fourth multimodal data is input into the marketing decision model to obtain the marketing decision corresponding to the fourth multimodal data.

7. The method according to claim 1, characterized in that, The method further includes: Acquire fifth multimodal data; wherein, the fifth multimodal data includes business department requirements, historical resource usage data, risk classification data, external market signals, and real-time performance indicators; The fifth multimodal data is input into the intelligent resource scheduling model to obtain the resource scheduling decision corresponding to the fifth multimodal data.

8. The method according to claim 1, characterized in that, The method further includes: Acquire sixth-modal data; wherein, the sixth-modal data includes bank statements, news text analysis, social media sentiment, macroeconomic indicators, industry reports, and regulatory policy documents; The sixth multimodal data is input into the capital flow prediction model to obtain the resource scheduling decision corresponding to the sixth multimodal data.

9. A bank transaction risk assessment application system, characterized in that, The bank transaction risk assessment application system includes a processor and a memory connected to the processor; the memory is used to store a computer program, which, when executed by the processor, is used to implement the method as described in any one of claims 1-8.

Images