Intelligent decision support method and system fusing multi-source financial data and security fence

Abstract

The invention discloses an intelligent decision support method and system fusing multi-source financial data and a security fence, and relates to the field of financial data processing, and the method comprises the steps: analyzing a customer portrait generation request, determining a portrait task scene, and converting the portrait task scene into a task context vector; according to the target customer identifier, recalling the original data, and decomposing the original data into a plurality of feature atoms; labeling metadata for each feature atom; constructing a financial feature topology network based on the plurality of feature atoms; calculating a dynamic confidence score of each feature atom; according to the dynamic confidence score, distributing the plurality of feature atoms to different confidence regions of a structured cue word template to generate a structured cue word; inputting the structured cue word into the large language model to generate a customer portrait report; and based on the customer portrait report, customer insight analysis and personalized verbal skills are generated and displayed. According to the method, effective credibility evaluation can be carried out on the multi-source financial data, so that the fact consistency is ensured.

Description

Technical Field

[0001] This application relates to the field of financial data processing, and in particular to an intelligent decision support method and system that integrates multi-source financial data with security fences. Background Technology

[0002] With the deepening development of fintech, commercial banks have accumulated massive amounts of customer data. This data comes from a wide range of sources and takes various forms, including structured transaction records, account information, and credit records; semi-structured customer service records and marketing campaign feedback; and unstructured customer interview texts and social media information. This multi-source, heterogeneous data provides a rich information foundation for building accurate customer profiles, but it also brings challenges such as inconsistent data quality and difficulty in assessing credibility. Different data sources exhibit significant differences in timeliness, completeness, and accuracy. For example, real-time transaction data has high credibility, while behavioral data collected by third parties may contain noise or bias.

[0003] In recent years, large language models (MLMs) have been widely applied in scenarios such as intelligent customer service, risk assessment, and personalized marketing in the financial sector due to their powerful semantic understanding and generation capabilities. In customer profile generation tasks, the traditional approach involves simply concatenating or formatting collected multi-source data before directly inputting it into an LLM for inference and generation. While this method can quickly produce customer profile reports in natural language, it suffers from a fundamental technical flaw: large language models themselves lack the ability to discern the quality and credibility of input data, failing to autonomously identify which information constitutes high-quality core facts and which are low-credibility auxiliary references.

[0004] In existing technologies, some solutions use rule engines or classification models to perform post-processing review of generated content, filtering out obvious violations or sensitive information. However, this post-processing review method can only address surface-level compliance issues and cannot solve deeper defects in the generation process. When a large language model receives input mixed with high- and low-quality information, it may make incorrect judgments when faced with conflicting information, ultimately leading to factual biases, logical contradictions, or even content completely detached from real data in the generated customer profile. This problem is particularly serious in financial scenarios, potentially leading to incorrect marketing decisions or misjudgments of risk, resulting in economic losses and a crisis of customer trust.

[0005] In summary, existing technologies have the drawback of being unable to effectively assess the credibility of multi-source financial data during the data input stage, resulting in a lack of ability to distinguish the strength of evidence during the generation of large language models, thus leading to issues of factual consistency. Summary of the Invention

[0006] This application provides an intelligent decision support method and system that integrates multi-source financial data and security fences, used to effectively assess the credibility of multi-source financial data in order to effectively ensure the consistency of facts.

[0007] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: Firstly, an intelligent decision support method integrating multi-source financial data and security fences is provided, the method comprising: In response to receiving a customer profile generation request, the customer profile generation request is parsed, the profile task scenario is determined, and the profile task scenario is converted into a task context vector, wherein the task context vector includes the target customer identifier; Based on the target customer identifier, the original data is retrieved from multiple financial data sources, and the original data is decomposed into multiple feature atoms, wherein each feature atom is an indivisible smallest unit of information. Label metadata for each of the aforementioned feature atoms; A financial feature topology network is constructed based on multiple feature atoms, with the feature atoms as nodes, and connections are established between nodes according to financial knowledge graphs and data relevance. A dynamic confidence score is calculated for each feature atom on the financial feature topology network, and the dynamic confidence score is dynamically adjusted based on the task context vector; Based on the dynamic confidence score, the multiple feature atoms are assigned to different confidence regions of the structured prompt word template to generate structured prompt words. The structured prompt word template includes a high-confidence fact region, a medium-confidence inference region, and a low-confidence reference region. The structured prompts are input into a large language model for inference to generate a customer profile report; Perform a traceability consistency audit on the customer profile report to verify whether the key conclusions in the customer profile report can be traced back to the feature atoms of the high-confidence fact area; Based on customer profile reports that have passed the traceability consistency audit, customer insight analysis and personalized communication scripts are generated and displayed through the account manager's workbench.

[0008] In one possible implementation of the first aspect, recalling original data based on multi-source financial data sources according to the target customer identifier includes: Based on the target customer identifier, internal business data is extracted from a pre-set internal tagging system. The internal business data includes wealth data, credit data, and payment data. Obtain anonymized external data from third-party data sources through a privacy-preserving computing interface; The internal business data and the external data are fused to retrieve the original data.

[0009] In another possible implementation of the first aspect, a financial feature topology network is constructed based on multiple said feature atoms, including: Multiple of the aforementioned feature atoms are used as network nodes; Extract product relationships and behavioral causal relationships from a pre-defined financial knowledge graph; Determine the temporal proximity and logical consistency among the characteristic atoms; Based on the product association, the behavioral causal relationship, and the temporal proximity and logical consistency between the feature atoms, edge connections are established between the network nodes; The financial feature topology network is generated based on the network nodes and the edge connections.

[0010] In another possible implementation of the first aspect, calculating a dynamic confidence score for each of the feature atoms on the financial feature topology network includes: A preset feature vector centrality algorithm is used to calculate the static centrality score of each feature atom on the financial feature topology network. The static centrality score reflects the objective importance of the feature atom in the financial feature topology network. The task context vector is applied to the financial feature topology network to adjust the weights of the edge connections related to the portrait task scenario, resulting in the adjusted financial feature topology network. The centrality score of each feature atom is recalculated on the adjusted financial feature topology network to obtain the dynamic confidence score of each feature atom.

[0011] In another possible implementation of the first aspect, the task context vector is applied to the financial feature topology network to adjust the weights of edge connections related to the profiling task scenario, including: Identify edge connections in the financial feature topology network that are semantically related to the portrait task scenario; When the profiling task scenario is marketing response evaluation, the weights of edge connections related to consumer behavior and interest preferences are dynamically increased. When the portrait task scenario is credit risk assessment, the weight of edge connections related to debt information, cash flow, and repayment history is dynamically increased.

[0012] In another possible implementation of the first aspect, the plurality of feature atoms are assigned to different confidence regions of the structured cue word template based on the dynamic confidence score, including: Feature atoms with dynamic confidence scores greater than the high confidence threshold are assigned to the high confidence fact area, and a first instruction identifier is added to the high confidence fact area. The first instruction identifier is used to instruct the large language model to use the feature atoms in the high confidence fact area as core factual evidence. Feature atoms whose dynamic confidence scores fall between the low confidence threshold and the high confidence threshold are assigned to the medium confidence inference region, and a second instruction identifier is added to the medium confidence inference region. The second instruction identifier is used to instruct the large language model to analyze the feature atoms in the medium confidence inference region as important clues. Feature atoms with dynamic confidence scores less than the low confidence threshold are assigned to the low confidence reference region, and a third instruction identifier is added to the low confidence reference region. The third instruction identifier instructs the large language model to use the feature atoms in the low confidence reference region as background reference information.

[0013] In another possible implementation of the first aspect, performing a traceability consistency audit on the customer profile report includes: Extract key conclusion statements from the customer profile report; Perform semantic analysis on each of the key conclusion statements to determine the core argument of each key conclusion statement; Retrieve supporting evidence related to the core argument from the characteristic atoms of the high-confidence fact region; If the supporting evidence is found, the key conclusion statement is marked as a high-confidence conclusion. If the supporting evidence is not found, the key conclusion statement is marked as a low-confidence inference, and the key conclusion statement is removed from the customer profile report and / or an uncertainty label is added.

[0014] In another possible implementation of the first aspect, customer insight analysis and personalized communication messages are generated based on customer profile reports that have passed traceability consistency audits, including: Extract customer attribute characteristics, behavioral trends, and potential needs from customer profile reports that have passed the traceability consistency audit; Based on the customer attribute characteristics, behavioral trends, and potential needs, customer insight analysis is generated. The system receives customer insight analysis and combines it with an intent recognition model, an intelligent agent decision-making model, and a financial knowledge base to generate personalized dialogue.

[0015] Secondly, this application provides an intelligent decision support system that integrates multi-source financial data and security fences, including: The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the aforementioned intelligent decision support method that integrates multi-source financial data and security fences.

[0016] Thirdly, this application provides a machine-readable storage medium storing instructions that cause a machine to execute the aforementioned intelligent decision support method integrating multi-source financial data and security fences.

[0017] The above technical solutions effectively solve the factual consistency problem caused by the inability of large language models to judge the quality of input data in existing technologies. By decomposing multi-source heterogeneous data into feature atoms and labeling metadata, fine-grained data management and quality tracking are achieved. The dynamic confidence calculation mechanism based on financial feature topology networks can comprehensively consider multiple factors such as data source, timeliness, network location, and task scenario to accurately assess the credibility of each information unit, significantly improving accuracy compared to traditional static rule evaluation methods. The design of structured prompt word templates, through clear confidence layering and instruction identification, effectively guides large language models to reason based on high-quality evidence, fundamentally reducing the risk of factual bias in generated content. The source consistency review mechanism establishes a complete traceability chain from conclusion to evidence, ensuring that each key judgment is fully supported by facts, significantly improving the credibility of customer profile reports. The final generated customer insight analysis and personalized scripts can directly guide business actions, effectively improving marketing conversion rates and customer satisfaction, and providing reliable technical support for the digital transformation of financial institutions.

[0018] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0019] Figure 1 A flowchart illustrating an intelligent decision support method that integrates multi-source financial data and security fences, provided as an embodiment of this application; Figure 2 A schematic diagram illustrating the construction process of a financial feature topology network provided in this application embodiment; Figure 3 This is a schematic diagram illustrating a process for performing a traceability consistency audit on a customer profile report, as provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. 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.

[0021] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0022] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0023] Figure 1 The illustration shows a flowchart of an intelligent decision support method that integrates multi-source financial data and security fences according to an embodiment of this application. Figure 1 As shown in the figure, this application provides an intelligent decision support method that integrates multi-source financial data and security fences. The method may include the following steps.

[0024] S110. In response to receiving a customer profile generation request, parse the customer profile generation request, determine the profile task scenario, and convert the profile task scenario into a task context vector, wherein the task context vector includes the target customer identifier. S120. Based on the target customer identifier, recall the original data from multiple financial data sources and decompose the original data into multiple feature atoms, where each feature atom is the smallest indivisible information unit. S130, Label metadata for each feature atom; S140. Construct a financial feature topology network based on multiple feature atoms. The financial feature topology network uses feature atoms as nodes and establishes connections between nodes based on financial knowledge graphs and data relevance. S150. Calculate the dynamic confidence score of each feature atom on the financial feature topology network. The dynamic confidence score is dynamically adjusted based on the task context vector. S160. Based on the dynamic confidence score, multiple feature atoms are assigned to different confidence regions of the structured prompt word template to generate structured prompt words. The structured prompt word template includes a high-confidence fact region, a medium-confidence inference region, and a low-confidence reference region. S170. Input the structured prompts into the large language model for reasoning and generate a customer profile report; S180. Perform a traceability consistency audit on the customer profile report to verify whether the key conclusions in the customer profile report can be traced back to the characteristic atoms of the high-confidence fact area; S190. Based on the customer profile report that has passed the traceability consistency audit, generate customer insight analysis and personalized scripts, and display them through the customer manager's workbench.

[0025] In this embodiment, the metadata includes the data source, timestamp, data type, and initial trust rating.

[0026] In addition to displaying the information through the account manager's workbench, the methods also include: Collect interaction data between account managers and customers, including communication records, customer feedback, and business processing results; The interactive data is fed back to the industry tagging system to update the attribute and behavioral characteristics of the target customers. Based on the updated attribute and behavioral features, the financial feature topology network is reconstructed to achieve iterative optimization of customer profiles.

[0027] Upon receiving a customer profile generation request, the first step is to parse the request content to clarify the business intent and application scenario. Customer profile generation requests originate from proactive triggers in the account manager's workbench, batch requests from the marketing campaign platform, or real-time assessment instructions from the risk management system.

[0028] In this embodiment, the parsing process employs natural language understanding technology to identify key elements in the request, including customer identification information, a description of the profile's purpose, and timeliness requirements. Different profile scenarios have significantly different focuses regarding customer information; for example, marketing response assessment scenarios focus more on customer consumption preferences, product ownership, and historical marketing response rates, while credit risk assessment scenarios emphasize debt structure, repayment ability, and credit history.

[0029] To achieve the digital representation of scene information, a pre-trained scene encoder is used to transform the profiling task scene into a high-dimensional dense vector, namely the task context vector. This vector is trained through contrastive learning on a large-scale financial business corpus, enabling it to capture the semantic features and business logic of different scenes. The task context vector not only contains the encoding of the scene type but also integrates the target customer identifier, typically using a unique identifier such as a customer number, to ensure the accuracy of subsequent data retrieval. Through this step, business requests in natural language form are transformed into a machine-processable structured representation, laying the foundation for subsequent personalized data processing. Simultaneously, the task context vector plays a guiding role throughout the entire processing flow, ensuring that the generated customer profile highly aligns with actual business needs.

[0030] After identifying the target customer, it is necessary to retrieve all original data related to that customer from multiple heterogeneous data sources and break it down into the smallest units of information. The data retrieval process is divided into two stages: internal data extraction and external data acquisition.

[0031] Specifically, internal data extraction retrieves a full range of customer business data from the bank's internal tagging system. This includes wealth management data such as deposit balances, wealth management product holdings, and precious metal investment records; credit data such as loan types, credit limits, repayment records, and overdue status; and payment data such as transaction details, consumption categories, and payment channel preferences. This data is typically stored in a structured format in a relational database and can be efficiently retrieved through SQL queries.

[0032] External data acquisition involves secure interaction with third-party data sources through privacy-preserving computation interfaces. Federated learning or multi-party secure computation techniques are used to obtain anonymized external features without disclosing the original data. Examples of such features include credit scores provided by credit reporting agencies, consumer behavior tags provided by e-commerce platforms, and interest preference tags provided by social media platforms.

[0033] The acquired internal and external data undergo feature alignment and fusion to form a complete original dataset. Subsequently, feature atom decomposition is performed, breaking down complex data records into indivisible minimum information units. For example, a transaction record can be decomposed into multiple independent feature atoms such as transaction amount, transaction time, merchant category, and transaction location, each representing a single-dimensional objective fact. The decomposition process follows the principle of atomicity, ensuring that each feature atom is semantically complete and independent, without containing composite information that can be further decomposed. Based on this, massive amounts of heterogeneous data can be transformed into a standardized set of feature atoms, providing a unified data foundation for subsequent credibility assessment and network construction. Simultaneously, the fine-grained representation of feature atoms allows credibility assessment to be precise down to each specific information point, avoiding information loss caused by coarse-grained data processing.

[0034] After completing the feature atom decomposition, metadata describing its source and attributes needs to be attached to each feature atom to support subsequent credibility assessment and source tracing verification. Metadata annotation includes four core dimensions. Data source annotation records the system or collection channel from which the feature atom was generated, such as core banking systems, credit management systems, third-party credit reporting platforms, etc. Data from different sources has different levels of authority and reliability.

[0035] Timestamps mark the time of business occurrence or data collection corresponding to the characteristic atoms. Timeliness is an important factor in evaluating the value of data. Recent data usually reflects the current status of customers better than historical data.

[0036] Data type labeling distinguishes whether a feature atom is a transaction-type fact, a state-type attribute, or a behavioral feature. Different types of data have different roles in business analysis and different credibility assessment standards. The initial credibility rating is based on the historical quality performance of the data source and the benchmark score preset by business rules. For example, transaction data from the core banking system may receive a high initial rating, while behavioral data collected externally may receive a medium rating.

[0037] In this embodiment, the annotation process employs an automated rule engine combined with manual verification. For regular data sources, it automatically matches preset metadata templates; for newly accessed data sources or abnormal data, it triggers a manual review process. Metadata is appended to feature atoms in the form of structured fields, forming complete data units. Through this step, a complete identity profile and quality label are established for each feature atom, enabling subsequent confidence calculations to comprehensively consider the multidimensional attributes of the data. Simultaneously, it provides necessary tracking information for the source verification of customer profile reports, ensuring that every conclusion can be traced back to the specific data source and time of generation.

[0038] After obtaining the set of feature atoms with metadata, a network structure that reflects the relationships between features needs to be constructed to support confidence calculation based on graph algorithms. The financial feature topology network uses feature atoms as nodes, with edges connecting them to represent the relationships between nodes. Network construction consists of two phases: node initialization and edge connection establishment.

[0039] Specifically, node initialization maps each feature atom to a node in the network. Node attributes include the content of the feature atom, metadata, and an initial credibility rating. Edge connections are established based on three types of relationships: First, product relationships in the financial knowledge graph, such as the business relationship between holding a mortgage and purchasing home insurance; these relationships are extracted from a pre-built financial knowledge graph. Second, behavioral causal relationships, such as the causal chain between increased income and the purchase of financial products, and the logical relationship between increased consumer spending and the need for increased bank card limits; these relationships are mined from historical data using causal inference algorithms. Third, data-level relevance, including temporal proximity and logical consistency. Temporal proximity refers to the closeness of feature atoms in the time dimension; for example, multiple transactions occurring on the same day may reflect the same consumption scenario. Logical consistency refers to the semantic support or contradiction between feature atoms; for example, high income and high consumption characteristics corroborate each other, while high debt characteristics contradict low risk ratings.

[0040] In one embodiment of this invention, the weights of edge connections are assigned based on the strength of the association, with strong associations receiving higher weights and weak associations receiving lower weights. This step organizes discrete feature atoms into a network structure with rich semantic relationships, providing a topological foundation for subsequent centrality calculations. Simultaneously, the network structure can capture complex interactions between features, enabling confidence assessment to consider not only the quality of individual features but also their position and role within the overall information network.

[0041] After constructing the financial feature topology network, the importance of each feature atom in the network is calculated, and dynamic adjustments are made according to the specific task scenario to obtain the final confidence score. The calculation process consists of two stages: static centrality calculation and dynamic adjustment.

[0042] The static centrality calculation employs the eigenvector centrality algorithm, whose core idea is that the importance of a node depends not only on the number of its connections but also on the importance of the nodes it connects to. Specifically, it iterative calculations ensure that the centrality score of each node equals the weighted sum of the centrality scores of all its neighboring nodes, and this iterative process continues until the scores converge.

[0043] Eigenvector centrality can identify core nodes in a network. These nodes typically connect multiple high-quality feature atoms and occupy key positions in the information network. Static centrality scores reflect the objective importance of feature atoms in the overall data network, but do not consider the differences in requirements for specific task scenarios.

[0044] The dynamic adjustment phase applies the task context vector to the network, adjusting the weights of edge connections according to scenario requirements. For marketing response assessment scenarios, the weights of edge connections related to consumer behavior, product preferences, and marketing response history are dynamically increased, while the weights of edge connections related to credit risk are decreased. For credit risk assessment scenarios, the weights of edge connections related to debt information, repayment ability, and cash flow stability are increased, while the weights of edge connections related to consumer preferences are decreased. Weight adjustment is achieved by calculating the similarity between the task context vector and the semantic vector of the edge connection; higher similarity indicates a greater relevance of the edge connection to the task scenario, resulting in a larger weight increase. Feature vector centrality calculation is then re-performed on the adjusted network to obtain a dynamic confidence score for each feature atom. This step enables scenario-adaptive confidence assessment; the same feature atom may obtain different confidence scores in different task scenarios, ensuring that the generated customer profile accurately matches business needs. Furthermore, the network topology-based calculation method can discover implicit relationships between data, improving the accuracy of confidence assessment.

[0045] After obtaining the dynamic confidence score of each feature atom, it is organized into a structured prompt word input to guide the large language model in generating high-quality customer profile reports. The structured prompt word template includes three confidence regions, each corresponding to a feature atom with a different confidence level. The high-confidence fact region contains feature atoms whose dynamic confidence scores exceed a preset high threshold. These features typically come from authoritative data sources, are highly timely, and occupy a core position in the network, representing core facts about the customer. A first instruction marker is added to this region, explicitly instructing the large language model to use this information as the basis for core facts, and to prioritize reasoning based on these features when generating conclusions, without contradicting them.

[0046] The medium-confidence inference region contains feature atoms with dynamic confidence scores between the low and high thresholds. These features have some reference value but are less reliable than high-confidence features, and can serve as clues for auxiliary analysis. A second instruction marker is added to this region, instructing the large language model to use this information as an important clue for comprehensive analysis. It can be used for inference and hypothesis, but the uncertainty should be noted.

[0047] The low-confidence reference region contains feature atoms with dynamic confidence scores below a low threshold. These features may originate from non-authoritative sources, have poor timeliness, or be located at the edge of the network, serving only as background reference. A third instruction marker is added to this region, instructing the large language model to treat this information as background reference and not as the primary basis for key conclusions. The allocation process is automatic based on dynamic confidence scores, ensuring that information of different qualities has a clear hierarchical structure within the prompt words. Through this step, the originally mixed multi-source data is transformed into a structured input with clear quality stratification, enabling the large language model to clearly distinguish the strength of evidence during the generation process and prioritize reasoning based on high-quality information. This effectively avoids the interference of low-quality information on the generated results, fundamentally solving the problem in existing technologies where models cannot judge the quality of input data.

[0048] After the structured prompts are constructed, they are input into a large language model for reasoning and generation. The large language model receives the structured prompts as input context, and based on its pre-trained language understanding and generation capabilities, combined with the instruction identifiers and hierarchical information in the prompts, generates a customer profile report in natural language form.

[0049] During the generation process, the model prioritizes feature atoms in the high-confidence fact region, using them as core arguments to construct a basic customer profile framework, including key information such as the customer's financial situation, product holdings, and transaction behavior characteristics. For feature atoms in the medium-confidence inference region, the model combines high-confidence information for comprehensive analysis to infer the customer's potential needs, risk preferences, or behavioral trends, but will reflect appropriate uncertainty in the description. For feature atoms in the low-confidence reference region, the model only uses them as supplementary background information and does not use them as the main basis for key conclusions.

[0050] In this embodiment, the generated customer profile report includes multiple dimensions such as an overview of the customer's basic information, financial status analysis, product usage behavior, risk characteristic assessment, and marketing opportunity identification. Presented in a structured paragraph format, it facilitates quick understanding and use by account managers. This step transforms structured feature data into easily understandable natural language descriptions. Simultaneously, leveraging the semantic understanding capabilities of a large language model, cross-feature correlation analysis and in-depth insights are achieved. The generated report not only includes statements of objective facts but also fact-based reasoning and recommendations, providing valuable references for business decision-making.

[0051] After generating a customer profile report, it needs to undergo rigorous quality review to ensure that all key conclusions in the report are supported by sufficient factual evidence. This consistency review can be implemented using natural language processing techniques and logical reasoning algorithms.

[0052] Specifically, the process begins by extracting key conclusion statements from customer profile reports. These statements typically contain judgments about customer characteristics, inferences about customer needs, or recommendations for business strategies. The extraction process employs syntactic analysis and semantic role labeling techniques to identify sentences expressing conclusive judgments. Each key conclusion statement undergoes semantic parsing to determine its core argument, i.e., the main fact or judgment asserted by the conclusion. Subsequently, supporting evidence related to the core argument is retrieved from the feature atoms in the high-confidence fact region. This retrieval process uses semantic similarity matching, calculating the semantic vector distance between the conclusion statement and the feature atom. If the distance is less than a threshold, the feature atom is considered valid supporting evidence. If sufficient supporting evidence is found, the key conclusion statement is marked as a high-confidence conclusion and retained in the final report. If no supporting evidence is found or the supporting evidence is insufficient, the key conclusion statement is marked as a low-confidence inference. Depending on the stringency, the statement may be removed from the report or an uncertainty label such as "This inference is based on limited information and is for reference only" may be added after it. In summary, this establishes a complete traceability chain from conclusion to evidence, ensuring that every key judgment in the customer profiling report can be traced back to high-quality original data. This effectively prevents factual biases or unfounded inferences that may arise from large language models, significantly improving the credibility and business value of the report.

[0053] Based on customer profile reports that have passed the traceability and consistency review, further customer insight analysis and personalized communication scripts are generated for business applications and displayed through the account manager's workbench. Customer insight analysis is an in-depth interpretation and business conversion of the customer profile report. The extraction process first identifies customer attribute characteristics in the report, including static attributes such as age, occupation, income level, and family structure, as well as dynamic attributes such as product holdings, transaction frequency, and risk preference. Then, it analyzes customer behavioral trends, identifying patterns of change in customer behavior through time series analysis, such as growth trends in consumer spending, changes in financial investment activity, and cyclical characteristics of credit demand. Based on attribute characteristics and behavioral trends, combined with business rules and machine learning models, it infers potential customer needs, such as housing loan needs, children's education savings needs, and retirement planning needs. Customer insight analysis is presented in a structured report format, including customer value assessment, demand forecasting, and marketing opportunity ranking.

[0054] Personalized script generation transforms customer insights into directly usable communication scripts. This process takes customer insight analysis as input, combines it with an intent recognition model to determine the business objectives of the current communication scenario, invokes an intelligent agent decision-making model to select the optimal communication strategy, and extracts relevant product information, policy terms, and script templates from a financial knowledge base. Ultimately, it generates personalized scripts that match customer characteristics and business objectives. The script content includes a complete communication flow, such as an opening, probing questions, reasons for product recommendation, and objection handling solutions. The generated customer insight analysis and personalized scripts are displayed on the account manager's workbench. This step transforms data analysis results into decision support tools that can directly guide business actions, effectively improving the work efficiency and service quality of account managers.

[0055] This embodiment effectively solves the factual consistency problem caused by the inability of large language models to judge the quality of input data in existing technologies. By decomposing multi-source heterogeneous data into feature atoms and labeling metadata, fine-grained data management and quality tracking are achieved. The dynamic confidence calculation mechanism based on financial feature topology networks can comprehensively consider multiple factors such as data source, timeliness, network location, and task scenario to accurately assess the credibility of each information unit, significantly improving accuracy compared to traditional static rule evaluation methods. The design of structured prompt word templates, through clear confidence layering and instruction identification, effectively guides large language models to reason based on high-quality evidence, fundamentally reducing the risk of factual bias in generated content. The source consistency review mechanism establishes a complete traceability chain from conclusion to evidence, ensuring that each key judgment is fully supported by facts, significantly improving the credibility of customer profile reports. The final generated customer insight analysis and personalized scripts can directly guide business actions, effectively improving marketing conversion rates and customer satisfaction, and providing reliable technical support for the digital transformation of financial institutions.

[0056] In one embodiment of this example, retrieving original data based on a target customer identifier and multiple financial data sources includes the following steps: S210. Based on the target customer identifier, extract internal business data from the preset internal labeling system. The internal business data includes wealth data, credit data, and payment data. S220. Obtain anonymized external data from a third-party data source through a privacy computing interface; S230. Perform feature fusion on internal business data and external data to retrieve the original data.

[0057] After identifying the target customer, it is necessary to accurately retrieve all business data related to that customer from multiple data source systems. Among them, internal business data extraction is the first stage of data retrieval, which involves reading the customer's full historical information from a pre-set internal tagging system.

[0058] In this embodiment, the internal tagging system is a unified customer data platform built within the commercial bank, integrating data from multiple business systems such as the core banking system, credit management system, and payment and settlement system.

[0059] Specifically, the wealth data extraction process first locates customer records in the customer information table based on the target customer identifier, and then correlates and queries the balance sheet, product holdings table, and investment transaction table. Wealth data includes current deposit balances, time deposit amounts and maturity dates, details of wealth management product holdings including product name, purchase amount, expected rate of return, and maturity date, fund investment portfolios including fund code, number of shares held, current net asset value, and historical returns, precious metal investment records including the holdings and market value of gold, silver, and other precious metals, and insurance product information including policy number, coverage amount, payment period, and beneficiary information.

[0060] Credit data extraction retrieves all credit and credit information of customers from the credit management system, including loan amount, loan term, interest rate type, number of repayments and remaining principal for housing loans; vehicle information, loan balance and monthly payment amount for auto loans; credit limit, used credit limit, billing date, repayment date and repayment history for the most recent six months for consumer loans; loan purpose, approved limit, withdrawal history and repayment plan for consumer loans; and historical overdue records including overdue time, overdue amount, overdue days and processing results.

[0061] Payment data extraction obtains customer transaction information from the payment clearing system, including all transaction records within the past year. Each transaction record includes the transaction time accurate to the second, transaction amount, transaction type such as consumption, transfer, cash withdrawal, counterparty information, merchant category code for identifying consumption scenarios such as catering, shopping, medical care, transportation, and various transaction channels.

[0062] In one implementation of this embodiment, the extraction process employs batch query and index optimization techniques to ensure the reading of massive amounts of data is completed within seconds. Simultaneously, sensitive fields such as complete card numbers are anonymized, retaining only the necessary information for analysis. Through this step, all business footprints of the customer stored in the bank's internal system are obtained. This data possesses high reliability and completeness, forming the core data foundation for building customer profiles.

[0063] External data acquisition is the second stage of data retrieval. It involves obtaining customer information dimensions that cannot be covered by the bank's internal systems through secure interfaces established with third-party data sources. Specifically, the privacy-preserving computation interface employs privacy-preserving technologies such as federated learning, multi-party secure computation, or trusted execution environments to enable computation and feature extraction to be completed without the data leaving its domain.

[0064] Specifically, third-party data sources include credit reporting agencies, e-commerce platforms, social media platforms, telecom operators, and public utilities.

[0065] Credit reporting agencies provide customers with summaries of their credit reports, including credit scores, credit information from other financial institutions, historical inquiry records, and public information. This information is anonymized and does not contain specific business details from other institutions; it only provides aggregated statistical features.

[0066] E-commerce platforms provide customers with consumer behavior tags through privacy computing interfaces, including purchase frequency levels such as high frequency, medium frequency, and low frequency; purchase amount ranges such as high consumption, medium consumption, and low consumption; preferred product category tags such as digital products, clothing, footwear, home furnishings, and books and audio-visual products; as well as behavioral characteristics such as promotion sensitivity and brand preference. These tags are generated by machine learning models based on customers' historical behavior on the e-commerce platform and are returned in the form of tags rather than raw transaction data.

[0067] Social media platforms provide customers with interest preference tags and social activity metrics, including topics of interest such as financial investment, travel, education, and health, content interaction frequency, and social circle characteristics. This information is also provided in the form of aggregated features.

[0068] Telecom operators provide communication behavior characteristics, including average monthly call duration, number of text messages, data usage, and frequently used geographical locations. These characteristics reflect customers' lifestyles and activity ranges. Public utilities provide statistical characteristics of water, electricity, and gas payment records, such as payment timeliness and average monthly cost levels. This information can indirectly reflect customers' residential stability and quality of life.

[0069] During the external data acquisition process, the privacy-preserving computation interface ensures that the raw data does not leave the data source's security domain, transmitting only the calculated aggregated or encrypted features to the bank's system, effectively protecting customer privacy and the data source's business secrets. The acquired external data undergoes standardization processing, converting it into a unified feature format to prepare for subsequent integration. This step effectively expands the information dimensions of the customer profile, filling gaps in the coverage of the bank's internal data, and enabling the profile to more comprehensively reflect the customer's true situation.

[0070] After acquiring internal business data and external data, the two types of data need to be merged to form a unified raw dataset. Since the internal data is mostly structured numerical and categorical fields, while the external data is mostly label and rating features, feature fusion faces challenges such as heterogeneous data formats, feature semantic alignment, and conflict resolution.

[0071] Specifically, the format standardization process normalizes all numerical features, maps numerical values ​​of different dimensions to a unified range such as 0 to 1, converts categorical features into numerical representations through one-hot encoding or label encoding, and converts textual labels into vector representations through a pre-trained semantic encoder.

[0072] Feature semantic alignment is a core step in data fusion, aiming to identify and correlate features describing the same customer attributes in internal and external data. For example, the average monthly spending amount in internal data and the spending range labels provided by external e-commerce platforms describe a customer's spending power, and a mapping relationship needs to be established.

[0073] In this embodiment, semantic alignment adopts a feature matching algorithm based on knowledge graphs. A pre-constructed financial domain feature ontology defines the semantic relationships of various features. By calculating the semantic similarity between feature names and feature descriptions, alignable feature pairs are identified.

[0074] For alignable features, the fusion strategy is selected based on their data type: for numerical features, a weighted average method is used, assigning different weights according to the credibility of the data source, with internal data usually receiving higher weights; for categorical features, a voting mechanism is used, selecting the category with the highest frequency as the fusion result; for vector features, a concatenation or attention mechanism is used for fusion.

[0075] Conflict resolution involves handling contradictory information between internal and external data. For example, internal data may show that a customer is at a low risk level, while external credit data may show overdue records. In such cases, it is necessary to judge based on the timeliness and credibility of the data, prioritizing the data source with stronger timeliness and higher credibility. If necessary, the conflicting information should be retained and marked for subsequent manual review.

[0076] The fused original dataset is stored in a unified data structure, with each feature field containing metadata such as feature name, feature value, data source identifier, and timestamp. In summary, fragmented information from multiple heterogeneous data sources can be integrated into a complete customer data view, providing high-quality data input for subsequent feature atomic decomposition and confidence assessment. Simultaneously, the semantic alignment and conflict resolution mechanisms during feature fusion effectively improve data consistency and accuracy, reducing the risk of error propagation in subsequent processing stages.

[0077] This embodiment achieves panoramic collection and high-quality integration of customer information through the systematic retrieval and fusion of multi-source financial data. Internal business data extraction ensures the integrity and credibility of core financial information, covering the entire business footprint of customers within the bank. External data acquisition, through privacy-preserving computation technology, effectively expands the information dimensions while protecting privacy, filling the gaps in the coverage of internal data. The feature fusion mechanism, through semantic alignment and conflict resolution, transforms heterogeneous data into a unified, high-quality dataset, laying a solid foundation for subsequent intelligent analysis. Compared to traditional single-data source analysis, this method provides more comprehensive and accurate customer insights, significantly improving the accuracy of customer profiling and the effectiveness of business decisions. Simultaneously, the application of privacy-preserving computation technology ensures data security and compliance, providing reliable technical support for data-driven business innovation in financial institutions.

[0078] Reference Figure 2 In one embodiment of this example, constructing a financial feature topology network based on multiple feature atoms includes the following steps: S310, Use multiple feature atoms as network nodes; S320. Extract product relationships and behavioral causal relationships from a pre-set financial knowledge graph; S330, Determine the temporal proximity and logical consistency between characteristic atoms; S340. Based on product association, behavioral causal relationship, and temporal proximity and logical consistency between feature atoms, establish edge connections between network nodes. S350 generates a financial feature topology network based on network nodes and edge connections.

[0079] After completing the metadata annotation of the feature atoms, these discrete information units need to be organized into a topological graph with a network structure to support subsequent confidence calculation based on graph algorithms.

[0080] Network node initialization is the first step in constructing the financial feature topology network, mapping each feature atom to an independent node in the network. Each node not only contains the core content of the feature atom but also carries complete attribute information.

[0081] Specifically, the core content of a node includes the specific numerical or categorical values ​​of the feature. Node attribute information includes the metadata annotated in the aforementioned steps, namely, data source identifiers such as core banking systems and third-party credit reporting platforms, timestamps indicating the generation or collection time of the feature, data type identifiers such as transaction-related facts, status-related attributes, and behavioral features, and an initial credibility rating reflecting the historical quality performance of the data source.

[0082] In addition, nodes also contain semantic vector representations of features, which are encoded into high-dimensional dense vectors by a pre-trained financial domain language model for subsequent semantic similarity calculations.

[0083] In this embodiment, the node initialization process adopts a graph database storage structure, which supports efficient node query and relationship traversal operations. For large-scale customer data containing thousands or even tens of thousands of feature atoms, node initialization needs to be completed within milliseconds; therefore, batch insertion and index optimization techniques are employed.

[0084] Each node is assigned a unique node identifier, typically in UUID format, to ensure global uniqueness in a distributed environment. Node attribute information is stored in key-value pairs, supporting flexible attribute expansion and dynamic updates. This step transforms feature atoms from simple data records into basic units for network analysis, laying the structural foundation for subsequent relationship establishment and topology analysis. Simultaneously, the rich attribute information of nodes provides necessary data support for multi-dimensional network analysis.

[0085] A financial knowledge graph is a pre-built domain knowledge base that stores entities, concepts, and their relationships within financial transactions. Product relationship extraction involves identifying business rules and marketing strategies related to the products held by customers from the knowledge graph.

[0086] Specifically, product relationships include product portfolio relationships, such as the bundled sale of mortgages and home insurance, and the dependency relationship between wealth management products and fund accounts. These relationships reflect the synergistic effect of products in the business process.

[0087] Product upgrade relationships describe the hierarchical evolution path between products, such as the criteria for upgrading a regular customer to a VIP customer, or the guidance strategies for converting demand deposits into time deposits. Product exclusivity relationships identify product combinations that cannot be held or recommended simultaneously, such as multiple loan products of the same type, or investment portfolios with conflicting risk levels.

[0088] The extraction process first identifies the feature atoms corresponding to the products currently held by the customer, and then queries the related edges of these product nodes in the financial knowledge graph to filter out the relationships related to the products actually held by the customer.

[0089] Behavioral causality extraction involves identifying causal chains and logical dependencies between customer behaviors. Behavioral causality includes the causal relationship between income and consumption; for example, an increase in salary usually leads to increased consumer spending, and bonuses may trigger large-scale consumption or investment. This relationship is obtained through statistical analysis of historical data and causal inference algorithms.

[0090] The causal relationship between debt and repayment describes the temporal dependence of loan disbursement and repayment behavior, such as the fixed monthly repayment date and behavioral patterns on repayment days after a mortgage is disbursed. The causal relationship between investment and returns reflects the logical chain between purchasing financial products, receiving returns, and redemption. The causal relationship between risk events and behavioral changes identifies the impact of negative events such as delinquency and complaints on subsequent behavioral patterns, such as changes in repayment behavior after delinquency and decreased product usage frequency after complaints.

[0091] In one embodiment of this study, the extraction process employs a combination of rule-based pattern matching and machine learning-based relationship prediction. The rule engine handles explicit business logic, while the machine learning model discovers implicit causal patterns. Through this step, rich prior knowledge is acquired from the domain knowledge base. This knowledge provides business semantic guidance for subsequent edge connection establishment, ensuring that the constructed network structure not only reflects statistical correlations at the data level but also embodies the inherent logic and causal mechanisms of financial business.

[0092] Temporal proximity and logical consistency are two important dimensions for mining the relationships between feature atoms from the data itself. Among them, temporal proximity measures the degree of proximity of feature atoms in the time dimension, based on the assumption of temporal clustering effect, that is, events that occur at close time often have an inherent relationship.

[0093] Specifically, the temporal proximity calculation first extracts the timestamp of each feature atom: for transaction-type features, the transaction occurrence time is used; for state-type features, the state update time is used; and for attribute-type features, the data acquisition time is used. Then, the time interval between any two feature atoms is calculated; the smaller the time interval, the higher the temporal proximity.

[0094] To quantify temporal proximity, a time decay function is used, commonly in the form of an exponential decay function. In this function, the temporal proximity score decreases exponentially with increasing time intervals, and the decay rate is determined by the business scenario. For high-frequency transaction scenarios such as payment transactions, the decay rate is large, and only hourly or daily proximity is considered. For low-frequency business scenarios such as loan applications, the decay rate is small, and monthly or even quarterly proximity can be considered.

[0095] Feature atoms whose time proximity scores exceed a preset threshold are marked as time-related. Examples include multiple consumer transactions on the same day, financial purchases within a week after salary is deposited, and changes in account balances before and after the repayment date. These time-related feature atoms may reflect the same business scenario or behavioral pattern.

[0096] Logical consistency measures the semantic and numerical relationships between feature atoms. It is based on the principle of information consistency, which states that different features describing the same customer should be logically consistent. Logical consistency judgment is divided into two levels: semantic consistency and numerical consistency.

[0097] Semantic consistency is evaluated by calculating the semantic vector similarity of feature atoms. High similarity indicates that the two features are semantically related. For example, "monthly consumption of 50,000 yuan" and "strong consumption ability" are semantically consistent, and "holding a mortgage" and "owning property" are semantically mutually supportive.

[0098] Numerical consistency is determined through business rules and statistical models. For example, high income characteristics and high consumption characteristics are numerically consistent, low risk rating and no overdue record are logically consistent, while high debt characteristics and low risk rating are logically contradictory and need to be marked as conflicting.

[0099] Logical consistency calculation also considers the numerical constraints between features. For example, the remaining principal of a loan should decrease with the number of repayment periods. Feature atomic pairs that violate these constraints are marked as logical conflicts and require data verification or confidence reduction in subsequent processing. In summary, the inherent relationships between feature atoms can be mined from both temporal and logical dimensions. These relationships reflect the temporal patterns and semantic constraints of the data, providing a data-driven basis for constructing a network structure that conforms to business logic.

[0100] After acquiring information on product relationships, behavioral causal relationships, temporal proximity, and logical consistency, edge connections need to be established between network nodes to form a complete topology. The edge connection establishment process comprehensively considers the above four types of relationships and adopts a multi-source relationship fusion strategy.

[0101] For any two nodes, first determine whether there is a product association relationship. If the products involved in the feature atoms corresponding to the two nodes have an association edge in the financial knowledge graph, then establish an edge connection of product association type between the two nodes. The weight of the edge is assigned according to the association strength, with high weights for strong associations such as product bundling sales and low weights for weak associations such as product category similarity.

[0102] Next, it is determined whether a causal relationship exists between the behaviors described by the two feature atoms. If a causal chain exists in the knowledge graph, then an edge connection of the causal relationship type is established, with the edge pointing from cause to effect. The weight is assigned according to the strength of the causal relationship. Deterministic causal relationships, such as repayment date triggering repayment behavior, are given high weight, while probabilistic causal relationships, such as income increase leading to consumption growth, are given medium weight.

[0103] The temporal proximity is assessed again. If the temporal proximity scores of two feature atoms exceed a threshold, a temporally related edge connection is established, with the edge weight equal to the temporal proximity score; the closer the times, the higher the weight.

[0104] Finally, logical consistency is assessed. If two feature atoms are consistent or conflicting in semantics or numerical value, an edge connection of the logical relation type is established. Consistent relations are assigned positive weights to indicate mutual support, while conflicting relations are assigned negative weights to indicate mutual contradiction.

[0105] In this embodiment, for node pairs that simultaneously satisfy multiple association relationships, multiple edges of different types can be established, or the weights of multiple relationships can be fused to obtain a comprehensive weight. The fusion method adopts weighted summation, and the weight coefficients of different relationship types are set according to the importance of the business. For example, the weight coefficient of causal relationship is higher than that of time association, and the weight coefficient of product association is higher than that of semantic consistency.

[0106] Once the edge connections are established, a directed or undirected weighted network is formed. The edges in this network not only represent relationships between nodes, but their type and weight also characterize the nature and strength of these relationships. Based on this, the multidimensional relationships obtained from knowledge and data mining can be transformed into the network's topology, enabling subsequent graph algorithms to propagate and compute on this structure, thereby achieving a global evaluation of the importance of feature atoms.

[0107] After node initialization and edge connection establishment are completed, a complete financial feature topology network can be generated. The network generation process integrates the node set and edge set into a unified graph data structure, which is stored in the form of an adjacency list or adjacency matrix.

[0108] The adjacency list approach is suitable for sparse networks. It maintains a list for each node to store all its neighboring nodes and their corresponding edge weights, offering high query efficiency and low storage space. The adjacency matrix approach is suitable for dense networks or scenarios requiring fast access to any pair of nodes. It uses a two-dimensional matrix to store the connection relationships and weights between nodes. A zero matrix element indicates no connection between nodes, while non-zero values ​​represent edge weights.

[0109] The generated financial feature topology network has multiple semantic layers. At the node level, it contains the content and attribute information of feature atoms. At the edge level, it contains the type and strength information of the association relationship. At the network level, it reflects the overall structure of customer data and the information flow pattern.

[0110] The topological characteristics of a network can be quantitatively analyzed using graph theory metrics. For example, network density reflects the tightness of connections between nodes, clustering coefficient reflects the clustering characteristics of local areas, and average path length reflects the efficiency of information propagation.

[0111] For typical customer data, financial feature topology networks typically exhibit small-world network characteristics, meaning that most nodes can be interconnected via short paths, while several highly connected core nodes serve as information hubs. These core nodes often correspond to key customer characteristics, such as income level, major product holdings, and core transaction behaviors. These characteristics will receive high centrality scores in subsequent confidence level calculations.

[0112] The network may also exhibit a community structure, with different communities corresponding to different business dimensions of customer data, such as wealth management communities, credit business communities, and payment behavior communities. Nodes within a community are tightly connected, while connections between communities are relatively sparse. In summary, this completes the transformation from discrete feature atoms to a structured network, providing a complete topological foundation for subsequent dynamic confidence calculation based on graph algorithms. At the same time, the network structure itself contains rich customer information patterns, supporting diverse graph mining and analysis tasks.

[0113] This embodiment constructs a financial feature topology network to achieve structured organization and semantic expression of multi-source heterogeneous customer data. Network nodes carry complete information about feature atoms, network edges depict multi-dimensional relationships between features, and the network topology reflects the overall pattern of customer data. Compared to traditional planar data representation, the topology network can explicitly express the complex interactions between features, enabling subsequent confidence calculations to consider not only the quality of individual features but also their position and role in the information network, significantly improving the accuracy of the assessment. The network construction process integrates financial domain knowledge and data-driven association mining, ensuring that the network structure conforms to both business logic and data reality, providing an interpretable analytical foundation for intelligent decision-making. This method provides a new technical paradigm for financial data analysis, supporting deeper customer insights and more accurate business decisions.

[0114] In one embodiment of this example, the dynamic confidence score of each feature atom is calculated on the financial feature topology network, including the following steps: S410. Using a preset feature vector centrality algorithm, calculate the static centrality score of each feature atom on the financial feature topology network. The static centrality score reflects the objective importance of the feature atom in the financial feature topology network. S420. Apply the task context vector to the financial feature topology network to adjust the weights of the edge connections related to the portrait task scenario, and obtain the adjusted financial feature topology network. S430. Recalculate the centrality score of each feature atom on the adjusted financial feature topology network to obtain the dynamic confidence score of each feature atom.

[0115] After constructing the financial feature topology network, it is necessary to quantitatively evaluate the importance of each feature atom in the network. The eigenvector centrality algorithm is a classic method for evaluating node importance in graph theory. Its core idea is that the importance of a node depends not only on the number of its connections, but also on the importance of the nodes it connects to.

[0116] In this embodiment, the static centrality score is calculated using an iterative convergence method. First, all nodes are assigned the same initial centrality score, which is usually set to 1 or 1 / N, where N is the total number of nodes.

[0117] Then, the iterative calculation phase begins. In each iteration, the new centrality score of each node is equal to the sum of the products of the centrality scores of all its neighboring nodes in the previous iteration and their corresponding edge weights. Specifically, for node i, its centrality score in the (t+1)th iteration is equal to the sum of the centrality scores of all its connected nodes j in the tth iteration multiplied by their edge weights, divided by a normalization constant to ensure convergence of the scores.

[0118] The iterative process continues until the centrality scores of all nodes change by less than a preset threshold, typically set to 0.0001, or until the maximum number of iterations, such as 100, is reached. The converged centrality scores reflect the global importance of a node in the network. High-centrality nodes are usually located at the core of the network, connecting multiple other important nodes, and play a key role in information propagation and network structure.

[0119] In financial feature topology networks, feature atoms with high centrality are often the core features of customer profiles, such as income level, main asset allocation, and core transaction behavior. These features are not only of high quality themselves, but also strongly correlated with multiple other high-quality features, forming a chain of evidence that corroborates each other.

[0120] In one implementation of this embodiment, the computational complexity of the static centrality score is related to the network size and the number of edges. For networks containing thousands of nodes, sparse matrix optimization and parallel computing techniques can complete the computation within seconds. Through this step, an objective assessment of the importance of each feature atom in the overall data network is obtained, providing a benchmark reference for subsequent dynamic adjustments.

[0121] Static centrality scores reflect the objective importance of feature atoms, but they do not consider the differences in requirements across specific business scenarios. Since different profiling tasks have significantly different focuses on customer information, the network structure needs to be dynamically adjusted based on the task context.

[0122] The task context vector contains the semantic encoding of the portrait task scenario. This vector, generated by the scenario encoder in the aforementioned steps, is a high-dimensional dense vector that captures the business characteristics and analysis focus of the scenario.

[0123] The process of applying task context vectors to the financial feature topology network is essentially adjusting the weights of edge connections in the network according to scenario requirements, thereby strengthening the association of features related to the scenario and weakening the association of features unrelated to the scenario.

[0124] Specifically, the weight adjustment process first requires identifying edge connections in the network that are semantically related to the task scenario. The identification method is to calculate the semantic similarity between the feature atoms associated with the edge connection and the task context vector.

[0125] For each edge in the network, extract the semantic vectors of the feature atoms corresponding to the two nodes it connects, concatenate the two semantic vectors or perform average pooling to obtain the semantic representation vector of the edge, and then calculate the cosine similarity between this vector and the task context vector. The similarity value is between -1 and 1. The higher the similarity, the more relevant the edge connection is to the task scenario.

[0126] For marketing response assessment scenarios, the semantic space of the task context vector is closer to features such as consumer behavior, product preferences, and marketing history. Therefore, edges connecting these features will receive higher similarity scores. The system dynamically increases the weights of these edges, with the increase proportional to the similarity score. For example, the original weights are multiplied by an adjustment factor greater than 1, such as 1.5 or 2.0. At the same time, the weights of edges related to credit risk and debt information are reduced by multiplying the original weights by an adjustment factor less than 1, such as 0.5 or 0.3.

[0127] For credit risk assessment scenarios, the semantic space of the task context vector is closer to features such as debt structure, repayment history, and cash flow status. The system increases the edge weights connecting these features and reduces the edge weights related to consumption preferences and hobbies.

[0128] After weight adjustment, the network topology remains unchanged, but the edge weight distribution changes. Features highly relevant to the task scenario have stronger internal connections, forming denser information clusters, while features less relevant to the scenario have less influence in the network. Based on this, the network structure can be adaptively adjusted for the specific task scenario, allowing subsequent centrality calculations to reflect feature importance within that scenario, rather than general objective importance.

[0129] After obtaining the adjusted financial feature topology network, the feature vector centrality algorithm needs to be re-executed to calculate the centrality score of each feature atom under the new network structure, i.e., the dynamic confidence score.

[0130] The recalculation process uses the same iterative algorithm as the static centrality calculation, but because the edge weights have been adjusted according to the task scenario, the information propagation pattern during the iteration process changes. In the adjusted network, edges related to the task scenario have higher weights, and these edges pass on more centrality scores in the iterative calculation, resulting in higher dynamic confidence scores for the nodes connected to these edges. Conversely, the weights of edges less related to the task scenario are reduced, and the nodes connected to them receive fewer scores during iteration, resulting in a relatively lower final dynamic confidence score.

[0131] After iterative convergence, each feature atom obtains a dynamic confidence score, which comprehensively reflects the positional importance of the feature atom in the overall network and its business relevance in a specific task scenario.

[0132] Comparing static centrality scores and dynamic confidence scores reveals that some features that are of average importance in static networks may gain significantly higher dynamic scores in specific scenarios due to their high relevance to the scenario. For example, customer interests may not be core features in general scenarios, but they become key information for predicting customer willingness to respond in marketing response evaluation scenarios.

[0133] Conversely, some features with high centrality in static networks may have their dynamic confidence scores reduced if they are weakly associated with the current task scenario, thus avoiding over-reliance on these features when the scenario is mismatched.

[0134] The calculation of dynamic confidence scores achieves a dual optimization of confidence assessment, considering both the objective quality and network location of the data, as well as the actual needs of the business scenario, ensuring that the generated customer profile is highly consistent with the specific application scenario. In summary, this completes the transformation from static to dynamic assessment, providing accurate confidence metrics for subsequent structured prompt word generation, allowing features of different qualities and relevance to be reasonably assigned to their corresponding confidence levels.

[0135] This embodiment achieves accurate quantitative assessment of the importance of feature atoms through a dynamic confidence score calculation mechanism based on the eigenvector centrality algorithm. Static centrality calculation provides an objective importance benchmark based on network topology, while the dynamic adjustment mechanism adaptively optimizes the network structure according to task scenario requirements. The recalculated dynamic confidence score reflects both data quality and business relevance. Compared to traditional rule-based or single-dimensional quality assessments, this method comprehensively considers multiple factors such as data source, timeliness, network location, and scenario requirements, significantly improving the accuracy and applicability of confidence assessment. The dynamic confidence score provides a scientific basis for subsequent structured layering of prompt words, ensuring that the large language model can reason based on high-quality and scenario-relevant evidence, effectively improving the accuracy and business value of customer profile generation.

[0136] In one embodiment of this invention, the task context vector is applied to the financial feature topology network to adjust the weights of edge connections related to the portrait task scenario, including the following steps: S510. Identify edge connections in the financial feature topology network that are semantically related to the portrait task scenario. S520. In the case of marketing response evaluation in the profiling task scenario, dynamically increase the weight of edge connections related to consumer behavior and interest preferences. S530. In the case of credit risk assessment in the portrait task scenario, dynamically increase the weight of edge connections related to debt information, cash flow and repayment history.

[0137] In the context of marketing response evaluation in the profiling task, the weights of edge connections related to consumer behavior and interests are dynamically increased, including: Calculate the semantic similarity between the task context vector and each edge connection; Identify edge connections that connect consumption frequency feature atoms, consumption amount feature atoms, product preference feature atoms, and channel usage feature atoms; The marketing scenario weight enhancement coefficient is applied to the identified edge connections, and the value of the marketing scenario weight enhancement coefficient ranges from 1.5 to 3.0; Calculate the adjusted weight of each edge connection based on semantic similarity and marketing scenario weight enhancement coefficient; Update the weight values ​​of the corresponding edge connections in the financial feature topology network.

[0138] In the context of credit risk assessment for profiling tasks, the weights of edge connections related to debt information, cash flow, and repayment history are dynamically increased, including: Calculate the semantic similarity between the task context vector and each edge connection; Identify the edge connections connecting the debt ratio feature atom, the average monthly cash flow feature atom, the overdue record feature atom, and the asset-liability ratio feature atom; Risk scenario weight enhancement coefficients are applied to the identified edge connections, with the value of the risk scenario weight enhancement coefficient ranging from 2.0 to 4.0; Calculate the adjusted weights of each edge connection based on semantic similarity and risk scenario weight enhancement coefficient; Update the weight values ​​of the corresponding edge connections in the financial feature topology network, and add an additional safety factor adjustment to the edge connections that are directly related to default risk.

[0139] Before applying the task context vector to the financial feature topology network, it is first necessary to identify which edge connections in the network have semantic relevance to the current profiling task scenario. Identifying the semantic relevance of edge connections is a prerequisite for weight adjustment and directly affects the accuracy and effectiveness of subsequent adjustments.

[0140] In this embodiment, the recognition process employs semantic matching technology, which comprehensively considers the content semantics of the features associated with the edge connections and the requirements of the task scenario.

[0141] For each edge connection in the network, the semantic information of the feature atoms corresponding to the two connected nodes is first extracted, including the textual description of the feature name, feature type, and feature value. For example, an edge connecting node A "average monthly spending of 8,000 yuan" and node B "holds bank card A" contains multiple dimensions of semantic information such as spending power, payment method, and customer level.

[0142] The semantic vectors of two nodes are fused into a unified semantic representation vector for the edge through concatenation or attention weighting. This vector captures the business semantics associated with the edge connection. Then, the cosine similarity between the semantic representation vector of the edge and the task context vector is calculated. The cosine similarity is obtained by dividing the dot product of the two vectors by the product of their respective magnitudes. The value ranges from -1 to 1. The closer the similarity is to 1, the more related the semantics are; the closer it is to -1, the opposite the semantics are; and the closer it is to 0, the no obvious relationship is.

[0143] For marketing response assessment scenarios, the task context vector learns marketing-related semantic features during pre-training, including dimensions such as consumer behavior, product preferences, channel usage, and marketing response history. Therefore, edges connecting these types of features will achieve high similarity scores. For credit risk assessment scenarios, the task context vector focuses more on risk-related features such as debt information, repayment ability, cash flow stability, and historical default records. Corresponding edge connections will achieve high similarity scores.

[0144] The identification process sets a similarity threshold, typically between 0.3 and 0.5. Edges with similarity exceeding the threshold are marked as semantically relevant to the task scenario and require weight adjustment. Edges with similarity below the threshold retain their original weights or undergo slight inverse adjustments. This step precisely filters out a subset highly relevant to the current task scenario from thousands or even tens of thousands of edge connections, providing a clear target range for subsequent differentiated weight adjustments. This avoids noise interference that may arise from global adjustments, ensuring the targeted and effective nature of the weight adjustments.

[0145] When the user profile task is marketing response assessment, the system needs to focus on feature dimensions that can predict a customer's willingness to respond to marketing requests and their likelihood of purchase. Since the core objective of marketing response assessment is to identify high-potential customers and predict their acceptance of specific products or services, features related to consumer behavior and interests are of critical value.

[0146] Specifically, the weight enhancement process first calculates the semantic similarity between the task context vector and each edge connection. This calculation has been completed in the aforementioned recognition step, and the similarity value serves as the basic coefficient for weight adjustment.

[0147] Then, the edge connections connecting specific types of feature atoms are identified. These feature atoms include consumption frequency features such as average monthly consumption frequency, quarterly consumption number, and annual active days; consumption amount features such as average monthly consumption amount, average single consumption amount, and proportion of large consumption; product preference features such as preferred product categories, brand loyalty, and price sensitivity; and channel usage features such as online and offline consumption ratio, mobile terminal usage frequency, and self-service channel preference.

[0148] The identification process is achieved through feature type label and keyword matching. Each feature atom has its category labeled in the metadata. The system quickly locates the target feature based on the category label and then finds all edge connections that connect these features.

[0149] The identified edges are connected to marketing scenario weight enhancement coefficients, which are set according to the importance of features to marketing prediction, ranging from 1.5 to 3.0. Specifically, edges connecting to purchase frequency and purchase amount features receive higher coefficients, such as 2.5 to 3.0, because these features directly reflect customers' spending power and activity levels, and are strong predictors of marketing response. Edges connecting to product preference features receive medium coefficients, such as 2.0 to 2.5, as these features help match customer needs with product characteristics, improving marketing accuracy. Edges connecting to channel usage features receive relatively low coefficients, such as 1.5 to 2.0, as these features are mainly used to optimize marketing outreach rather than predict response intentions.

[0150] The adjusted weights are calculated by multiplying the original weight by the semantic similarity and then by the marketing scenario weight enhancement coefficient. This ensures that the adjustment reflects both the scenario relevance and the original association strength. For example, an edge with an original weight of 0.8, a semantic similarity of 0.7, and an enhancement coefficient of 2.5 will have an adjusted weight of 0.8 × 0.7 × 2.5 = 1.4.

[0151] After calculation, the weight values ​​of the corresponding edge connections in the financial feature topology network are updated, and the new weights are written to the edge attribute fields of the graph database, completing the dynamic adjustment of the network structure. Through this step, the association of key features in the marketing scenario is significantly strengthened. These features will transmit more importance scores in subsequent centrality calculations, enabling feature atoms related to marketing response to obtain higher dynamic confidence scores. This ensures that the generated customer profile can accurately capture the customer's marketing value and response potential, providing strong support for precision marketing.

[0152] When the profiling task is credit risk assessment, the system needs to focus on the feature dimensions that reflect a customer's repayment ability and default risk. Since the core objective of credit risk assessment is to determine a customer's creditworthiness and predict the likelihood of future defaults, features related to debt information, cash flow, and repayment history play a decisive role.

[0153] The weighting process also begins by calculating the semantic similarity between the task context vector and each edge connection, providing a quantitative basis for weight adjustment.

[0154] Then, the edges connecting risk-related characteristic atoms are identified. These characteristic atoms include debt ratio characteristics such as the ratio of total liabilities to total assets and the monthly payment to income ratio; monthly cash flow characteristics such as salary income, operating income, and other sources of income; overdue record characteristics such as the number of historical overdue payments, the longest overdue days, the overdue amount, and the overdue frequency; and asset-liability ratio characteristics such as the ratio of current assets to current liabilities, net asset size, and asset liquidity.

[0155] The identification process uses a risk feature tag library for matching. This library predefines all feature types and keywords related to credit risk. It quickly filters out risk-related features based on the metadata tags of feature atoms and finds the edges connecting these features.

[0156] The weight enhancement coefficient for risk scenarios applied to the identified edge connections ranges from 2.0 to 4.0, which is significantly higher than the enhancement coefficient for marketing scenarios, reflecting the higher requirements for the reliability of evidence in risk assessment.

[0157] Specifically, the side connecting to overdue payment history characteristics receives the highest coefficient, such as 3.5 to 4.0, because historical default behavior is the strongest predictor of future default risk, and these characteristics have extremely high risk indicative value. The side connecting to debt ratio and debt-to-equity ratio characteristics receives a relatively high coefficient, such as 3.0 to 3.5. These characteristics directly reflect a customer's debt burden and repayment ability, and are core indicators for risk assessment. The side connecting to average monthly cash flow characteristics receives a moderate coefficient, such as 2.5 to 3.0. Cash flow stability is an important guarantee of repayment ability, but it needs to be considered in conjunction with other characteristics for comprehensive judgment.

[0158] The adjusted weights are calculated similarly to those in the marketing scenario, using the product of the original weights, semantic similarity, and the risk scenario weight enhancement coefficient. Furthermore, for edge connections directly related to default risk, the system adds an additional safety coefficient adjustment, typically set to 1.2 to 1.5, to further increase the weights of these key risk features.

[0159] The introduction of a safety factor reflects the principle of prudence in risk management, giving higher weight to risk factors that may lead to serious consequences, and ensuring that risk assessments do not overlook key risk signals.

[0160] After the weights are updated, risk-related features form high-weight subnetwork structures within the network. These features dominate the information propagation process during centrality calculation, resulting in significantly improved dynamic confidence scores for risk feature atoms. This step effectively strengthens key risk features in credit risk assessment scenarios, ensuring that the generated customer profiles accurately identify high-risk customers and potential default signals. This provides a reliable basis for credit approval, credit limit management, and risk warning, effectively reducing the credit risk exposure of financial institutions.

[0161] The weighting adjustment mechanism is designed to fully consider the specific needs and risk preferences of different business scenarios. There are significant differences in the weighting enhancement coefficients between marketing and risk scenarios, reflecting the fundamental difference in the severity of the decision consequences between the two types of scenarios.

[0162] Decision-making errors in marketing scenarios typically lead to wasted marketing costs or lost opportunities, with relatively controllable consequences. Therefore, the weighting enhancement coefficient is set relatively mildly, allowing for a certain degree of uncertainty and exploration. In contrast, decision-making errors in risk scenarios can lead to serious consequences such as loan defaults and financial losses, thus requiring a more aggressive weighting enhancement coefficient.

[0163] In addition, the introduction of an additional security factor provides a double safeguard for risk assessment, ensuring that critical risk signals are not diluted or ignored due to the complexity of the network structure.

[0164] The weight adjustment process also considers the interaction effect between features. For edges that connect multiple high-value features at the same time, the weight increase will be more significant due to the cumulative effect of semantic similarity, forming a positive feedback mechanism for reinforcement learning.

[0165] The adjusted network structure not only alters the edge weight distribution but also implicitly changes the path and speed of information flow. This transforms task-relevant feature subnetworks into high-speed channels for information propagation, while irrelevant features are marginalized. This structural change provides a topological foundation for subsequent centrality calculations, optimizing the network structure for specific scenarios. In summary, this enables precise control of the network structure and scenario adaptation, allowing the same set of feature data to generate differentiated confidence assessment results based on different business needs, significantly improving the flexibility and applicability of intelligent decision support systems.

[0166] This embodiment achieves scenario-adaptive optimization of the financial feature topology network through a dynamic weight adjustment mechanism based on task context. Semantic relevance recognition ensures the accuracy and targeting of weight adjustments, avoiding noise interference from global adjustments. Differentiated weight enhancement strategies for marketing and risk scenarios fully reflect the specific needs of different business scenarios: marketing scenarios focus on customer value mining and response prediction, while risk scenarios emphasize risk signal identification and prudent assessment. The scientific setting of weight enhancement coefficients and security coefficients ensures that key features are sufficiently strengthened while maintaining the rationality and interpretability of the network structure. The adjusted network provides a scenario-optimized topological foundation for subsequent dynamic confidence calculations, enabling the generated customer profiles to accurately match business needs and significantly improving the accuracy and business value of intelligent decision support. This method provides a unified technical framework for multi-scenario applications in the financial field, supporting flexible scenario switching and personalized decision support, effectively improving the digital operation capabilities and risk management level of financial institutions.

[0167] In one embodiment of this example, multiple feature atoms are assigned to different confidence regions of the structured cue word template based on dynamic confidence scores, including the following steps: S610. Assign feature atoms with dynamic confidence scores greater than the high confidence threshold to the high confidence fact area, and add a first instruction identifier to the high confidence fact area. The first instruction identifier is used to instruct the large language model to use the feature atoms in the high confidence fact area as the core fact basis. S620. Assign feature atoms with dynamic confidence scores between the low confidence threshold and the high confidence threshold to the medium confidence inference region, and add a second instruction identifier to the medium confidence inference region. The second instruction identifier is used to instruct the large language model to analyze the feature atoms in the medium confidence inference region as important clues. S630. Assign feature atoms with dynamic confidence scores less than the low confidence threshold to the low confidence reference region, and add a third instruction identifier to the low confidence reference region. The third instruction identifier instructs the large language model to use the feature atoms in the low confidence reference region as background reference information.

[0168] After obtaining the dynamic confidence score of each feature atom, the feature atom needs to be assigned to different regions of the structured cue word template according to the score, so as to achieve a hierarchical expression of evidence strength.

[0169] The first step is the allocation of the high-confidence fact zone, which contains the most reliable and important customer information. The allocation process begins by setting a high-confidence threshold, which is dynamically determined based on the business scenario and data quality characteristics. It is typically set between the 75th and 85th percentiles of the dynamic confidence score distribution to ensure that approximately 15% to 25% of high-quality features are included in this zone.

[0170] For scenarios with high reliability requirements, such as risk assessment, the threshold can be set higher, such as the 85th percentile, to rigorously screen core evidence. For exploratory scenarios, such as marketing analysis, the threshold can be appropriately lowered, such as the 75th percentile, to retain more valuable information.

[0171] Iterate through all feature atoms and compare their dynamic confidence scores with a high confidence threshold. Feature atoms with scores exceeding the threshold are assigned to the high-confidence fact region. These features typically have the following characteristics: they come from authoritative data sources such as core banking systems; they are highly timely, such as recent transaction data; they occupy a core position in the network, such as high-centrality nodes; and they are highly relevant to the task scenario, such as overdue records in risk scenarios.

[0172] After allocation, a first instruction identifier is added to the high-confidence fact area. This identifier explicitly indicates the processing strategy of the large language model in natural language form.

[0173] In this embodiment, the typical first instruction identifier content is "The following information is a core fact that has been rigorously verified. Please use this information as the main basis for analysis. All key conclusions generated must be based on these facts and must not contradict them."

[0174] The instruction markers guide the model to prioritize information in the relevant area through emphatic words such as "core facts", "main basis" and "must be based on", and prevent the model from generating content that conflicts with high-confidence facts through restrictive statements such as "must not contradict".

[0175] The first instruction identifier can also include formatting requirements. This step presents the most reliable evidence and clarifies its authority, establishing a solid factual foundation for the reasoning process of the large language model and ensuring that the core content of the generated customer profile has high credibility and strong evidentiary support.

[0176] The allocation of the medium-confidence inference region handles information that has some reference value but is less reliable than high-confidence features. The allocation process sets two boundaries: a low-confidence threshold and a high-confidence threshold. The low-confidence threshold is usually set to the 40th to 50th quantile of the dynamic confidence score distribution to ensure that the proportion of intermediate-level features is about 25% to 35%.

[0177] The system filters feature atoms whose dynamic confidence scores fall between two thresholds. These features may come from data sources with moderate confidence, such as third-party platforms, or historical data with moderate timeliness, or be in a secondary core position in the network, or have some correlation with the task scenario but are not directly related.

[0178] After the allocation is completed, add a second instruction label to the area. The typical content is: "The following information has certain reference value and can be used as an important clue for comprehensive analysis and inference. However, please note the uncertainty of this information. When drawing conclusions based on this information, cautious expressions such as 'maybe', 'tend to', and 'preliminary judgment' should be used."

[0179] The second instruction label encourages the model to use this information for in-depth reasoning through terms such as "important clues" and "comprehensive analysis," while it prevents the model from making definitive judgments by over-reliance on moderate-quality evidence through constraints such as "pay attention to uncertainty" and "cautious expression."

[0180] The instructions can also require the model to cross-validate with high-confidence facts when using medium-confidence information. This step provides the large language model with information resources to aid analysis, supporting deeper customer insights and behavioral predictions. At the same time, explicit uncertainty cues prevent inferences from being mistaken for definitive facts, maintaining a balance between accuracy and credibility of the generated content.

[0181] The assignment of low-confidence reference regions handles features with low confidence but which may provide contextual information. The assignment process assigns feature atoms with dynamic confidence scores below the low-confidence threshold to this region. These features typically come from non-authoritative data sources, have poor timeliness, are located at the edge of the network, or have weak relevance to the task scenario.

[0182] After allocation, add a third instruction label to the low confidence reference area. The typical content is: "The following information is for background reference only and should not be used as the main basis for key conclusions. It can be used to enrich the completeness of the customer profile, but it should be clearly marked as 'reference information' or 'information to be verified' when it is described, and should not be presented as a certain fact."

[0183] The third instruction label explicitly limits the scope of information use in this area through expressions such as "for background reference only" and "should not be used as the primary basis," preventing low-quality information from interfering with core judgments. At the same time, the expression "rich and complete" encourages models to appropriately use this information to supplement profile details, but must be accompanied by annotations such as "reference information" and "to be verified" to ensure that information users can distinguish between different quality levels of content.

[0184] The instructions can also require the model to make explicit disclaimers when using low-confidence information. This step achieves the reasonable use and risk isolation of low-quality information, neither completely discarding its potential value nor allowing it to mislead decision-making through clear labeling and constraints, thus maintaining a balance between the comprehensiveness and reliability of the customer profile.

[0185] The three-layer partitioning design of the structured cue template reflects the scientific principle of evidence hierarchical management. The characteristic atoms in different confidence regions form a clear hierarchical structure in terms of quantity distribution, quality characteristics and usage constraints.

[0186] Specifically, the high-confidence fact area typically contains 15% to 25% of the characteristic atoms, determining the basic direction and key conclusions of the profile. The medium-confidence inference area contains 25% to 35% of the characteristic atoms, providing rich analytical clues and inference basis, making the profile more comprehensive and in-depth. The low-confidence reference area contains 40% to 60% of the characteristic atoms, providing completeness and diversity, but does not affect the core judgment.

[0187] The threshold settings for the three-tier partitioning can be dynamically adjusted based on specific business needs and data quality distribution. For scenarios with generally high data quality, the thresholds for each tier can be increased to rigorously filter features from each region. For scenarios with inconsistent data quality, the thresholds can be appropriately lowered to ensure that each region has sufficient information to support analysis.

[0188] The instruction labeling uses natural language rather than code instructions, enabling large language models to accurately grasp usage requirements through their language understanding capabilities. Compared to simple weight annotations or priority markers, natural language instructions can convey richer and more detailed usage strategies, including multi-dimensional constraints such as usage scenarios, expression methods, and validation requirements. In summary, this completes the transformation from numerical confidence scores to structured prompt word input, providing large language models with a clear hierarchy of evidence and explicit usage guidance, fundamentally solving the problem in existing technologies where models cannot distinguish the quality of input data.

[0189] This embodiment achieves scientific hierarchical and structured organization of multi-source heterogeneous features through a three-layer partitioning mechanism based on dynamic confidence scores. The high-confidence fact zone ensures the authoritative status of core evidence, the medium-confidence inference zone supports in-depth analysis and reasonable inference, and the low-confidence reference zone provides background supplementation and integrity assurance. The threshold settings and instruction identifier design of the three-layer partitioning fully consider business scenario requirements and data quality characteristics, achieving a balance between flexibility and standardization. Compared to traditional flat data input, this method can explicitly convey evidence strength information to large language models, guiding the model to reason based on high-quality evidence, effectively avoiding interference and misleading information from low-quality information, significantly improving the factual consistency and credibility of the generated content, and providing reliable technical support for intelligent decision support in the financial field.

[0190] Reference Figure 3 In one embodiment of this example, performing a source consistency audit on the customer profile report includes the following steps: S710. Extract key conclusion statements from customer profile reports; S720. Perform semantic analysis on each key conclusion statement to determine the core argument of the key conclusion statement; S730. Retrieve supporting evidence related to the core argument from the characteristic atoms in the high-confidence fact region; S740. When supporting evidence is found, mark key conclusion statements as high-confidence conclusions. S750. If no supporting evidence is found, mark key conclusion statements as low-confidence inferences and remove key conclusion statements from the customer profile report and / or add uncertainty annotations.

[0191] After generating customer profile reports using the large language model, the report content needs to undergo rigorous quality review to ensure that every key judgment is based on sufficient factual evidence.

[0192] Among them, extracting key conclusion statements is the first step in the traceability consistency audit, and the goal is to identify statements that express judgments, conclusions or recommendations from the report in natural language form.

[0193] In this embodiment, the extraction process employs natural language processing techniques, combining syntactic analysis and semantic role labeling. First, the customer profile report is segmented into sentences, breaking the report text down into independent sentence units. Then, syntactic analysis is performed on each sentence to identify its subject-verb-object structure and semantic roles.

[0194] Key concluding statements typically have the following linguistic features: they contain judgmental verbs such as "is," "has," "behaves as," and "belongs to"; they contain evaluative adjectives such as "high," "low," "strong," "weak," "excellent," and "good"; they contain concluding adverbs such as "therefore," "in summary," and "overall"; or they contain suggestive expressions such as "suggest," "should," and "can be considered."

[0195] The system identifies key conclusion statements by combining rule matching and machine learning classifiers. The rule engine processes sentences with obvious linguistic markers, while the classifier judges the semantic function of sentences based on pre-trained language models and identifies implicit conclusive expressions.

[0196] The extraction process also requires identifying conclusive clauses in complex sentences, such as "because the customer has stable income and no overdue records, the credit risk is low," where "low credit risk" is the key conclusion.

[0197] After extraction, the system stores all key conclusion statements and their locations in the report as a structured list, providing input for subsequent step-by-step review. This step accurately identifies the key judgments requiring verification from the complete customer profile report, avoiding ineffective review of descriptive content and improving review efficiency and focus.

[0198] After extracting key conclusion statements, each statement needs to undergo semantic analysis to clarify its core argument. The goal of semantic analysis is to transform natural language expressions into structured semantic representations, facilitating subsequent evidence retrieval and matching.

[0199] Specifically, the parsing process first identifies the subject entity of the conclusion statement, that is, the object described by the conclusion, usually the customer themselves or a certain attribute dimension of the customer such as "investment intention", "credit risk", or "spending ability". Then it identifies the assertion content of the conclusion, that is, the judgment or evaluation of the subject entity, such as "high", "low", "stable", or "active". Next, it identifies the limiting conditions of the conclusion, including the time range such as "the last three months" or "current", the degree modifiers such as "very" or "relatively", and the applicable scenario such as "in a marketing scenario".

[0200] Semantic parsing employs dependency parsing and semantic role labeling techniques. Dependency parsing identifies grammatical dependencies between words, determining the core verbs and argument structure of the sentence. Semantic role labeling identifies the semantic roles of verbs, such as agent, patient, and instrument, clarifying the various components of the argument.

[0201] The parsing results are stored in a structured format, including fields such as entity identifier, attribute type, assertion value, and limiting conditions. For complex conclusion statements, the parsing process also needs to identify causal relationships and logical connections. For example, in the statement "Because of stable income, this customer has strong repayment ability," it is necessary to parse out that "stable income" is a prerequisite, "strong repayment ability" is the core argument, and there is a causal relationship between the two.

[0202] The parsing process also transforms evaluative expressions in natural language into quantifiable semantic representations. For example, "higher" can be mapped to a numerical range [0.7, 0.9], and "very high" can be mapped to [0.9, 1.0], facilitating matching with the numerical information of feature atoms. In summary, ambiguous natural language conclusions can be transformed into precise semantic structures, providing clear query targets for subsequent evidence retrieval and ensuring that the review process can accurately determine the logical connection between conclusions and evidence.

[0203] After identifying the core argument, it is necessary to search for evidence supporting that argument within the characteristic atoms of the high-confidence fact region. Evidence retrieval is a core step in the consistency review process and directly determines the credibility of the conclusion.

[0204] The retrieval process employs a multimodal matching strategy, comprehensively considering semantic similarity, numerical consistency, and logical relevance.

[0205] First, semantic similarity matching is performed. The semantic representation vector of the core argument is compared with the semantic vector of each feature atom in the high-confidence fact region. The similarity is measured by cosine distance or Euclidean distance. Feature atoms with similarity exceeding a preset threshold, such as 0.6, are marked as semantically related.

[0206] Next, numerical consistency verification is performed. For arguments containing quantitative assertions, it is necessary to verify whether the values ​​of the feature atoms support the assertion. For example, for the argument "the average monthly consumption amount is relatively high", it is necessary to retrieve the feature atom of average monthly consumption amount and verify whether its value is indeed in a relatively high range. If the feature atom shows "average monthly consumption of 8,000 yuan" and this value is in the top 30 percentile among the customer group, then it is determined that the values ​​are consistent.

[0207] A further logical correlation analysis is performed. For arguments containing causal relationships or chains of reasoning, multiple feature atoms supporting the reasoning process need to be retrieved. The retrieval process also considers temporal consistency; the time constraints in the argument need to match the timestamps of the feature atoms. For example, if the argument involves behavior within the "last three months," only feature atoms with timestamps within the last three months will be retrieved.

[0208] The search results are returned as a list of evidence, containing all relevant feature atoms supporting the argument and their matching scores. Arguments for which no supporting evidence was found are marked as unsupported and require further processing in subsequent steps. This step establishes a traceability chain from conclusion to evidence, ensuring the verifiability and interpretability of the generated content and guaranteeing that every key judgment in the customer profile report has a solid factual basis.

[0209] After completing the evidence retrieval, it is necessary to mark and process the credibility of key conclusion statements based on the retrieval results.

[0210] For conclusions supported by sufficient evidence, the system marks them as high-confidence conclusions. These conclusions are retained in the final report, and evidence citations can be selectively added to enhance the report's credibility and transparency. High-confidence conclusions typically account for 60% to 80% of the report's key conclusions, reflecting the ability of the large language model to reason effectively based on high-quality evidence under the guidance of structured prompts.

[0211] For conclusions that lack supporting evidence, the system marks them as low-confidence inferences. These conclusions may be speculations made by the model based on information with medium to low confidence, or they may be over-reasoning and conjecture by the model.

[0212] The handling strategy includes two options: complete removal or adding uncertainty markers. The removal strategy is suitable for rigorous risk assessment scenarios, where any conclusions lacking high-confidence evidence are deleted to ensure the absolute reliability of the report, but this may result in a relatively brief report.

[0213] Uncertainty labeling strategies are suitable for exploratory analysis scenarios, preserving these inferential conclusions but adding explicit limiting statements.

[0214] The processing decisions can be configured according to business scenarios and risk preferences. Risk-sensitive scenarios adopt a removal strategy, while value mining scenarios adopt a labeling strategy.

[0215] The system also generates an audit report, statistically analyzing the number and proportion of high-credibility conclusions and low-credibility inferences, as well as a list of removed or flagged conclusions, allowing business personnel to evaluate report quality and model performance. In summary, this comprehensive quality audit of customer profile reports ensures that the content presented to users is highly credible and strongly supported by evidence, effectively preventing factual biases and unfounded inferences that may arise from large language models, and significantly improving the reliability and business value of the intelligent decision support system.

[0216] This embodiment achieves rigorous quality control over the content generated by the large language model through a source consistency audit mechanism. Key conclusion extraction precisely locates the judgmental content requiring verification; semantic parsing transforms natural language into a verifiable structured representation; evidence retrieval establishes a complete traceability chain from conclusion to evidence; and credibility labeling and processing ensure the high reliability of the final report. This method fundamentally solves the problem of factual consistency in the content generated by the large language model. Compared to traditional post-event rule audits, source consistency audits can delve into the evidentiary level of each key conclusion, performing precise logical verification and fact-checking. The introduction of this audit mechanism enables the intelligent decision support system to effectively control the quality risks of the generated content while fully leveraging the semantic understanding and generation capabilities of the large language model. This provides reliable technical support for key business decisions in the financial sector and significantly enhances the professionalism and authority of customer profile reports.

[0217] In one embodiment of this example, customer insight analysis and personalized communication scripts are generated based on customer profile reports that have passed the traceability consistency audit, including the following steps: S810. Extract customer attribute characteristics, behavioral trends, and potential needs from customer profile reports that have passed the traceability consistency audit; S820: Generate customer insight analysis based on customer attribute characteristics, behavioral trends, and potential needs; S830 receives customer insight analysis and combines it with intent recognition models, intelligent agent decision-making models, and financial knowledge bases to generate personalized scripts.

[0218] This includes combining an intent recognition model, an intelligent agent decision-making model, and a financial knowledge base to generate personalized dialogue, including: By analyzing customer communication history using an intent recognition model, we can identify the customer's current communication intent. Based on customer insight analysis and the customer's current communication intent, the recommendation strategy is determined through the intelligent agent decision-making model. Retrieve product information and sales script templates that match the recommendation strategy from a financial knowledge base; Based on the personalized information in the customer profile report, the dialogue template is customized and adjusted to generate personalized scripts.

[0219] After obtaining a high-quality customer profile report that has passed the traceability consistency audit, it is necessary to extract key information dimensions from the report that can directly support business decisions.

[0220] Among them, customer attribute feature extraction focuses on describing the static characteristics and identity tags of customers, including basic attributes such as age group, occupation type, income level, family structure, asset size, and risk preference.

[0221] In this embodiment, the extraction process uses named entity recognition and attribute extraction technology to identify statements and phrases in the report that express attribute information.

[0222] Behavioral trend extraction focuses on the temporal changes and development trends of customer behavior patterns, including dynamic characteristics such as the growth or decline trend of consumer spending, changes in financial investment activity, fluctuations in product usage frequency, and shifts in channel preferences. The extraction process uses time series analysis and trend recognition algorithms to identify statements from reports that contain time comparisons, descriptions of changes, and trend judgments.

[0223] Potential demand extraction uncovers unmet customer needs for financial services and potential future business opportunities, including loan needs, insurance needs, wealth management needs, and children's education savings needs. The extraction process combines a demand inference model and a business rule engine to identify demand signals expressed in the report.

[0224] The extraction process also involves structuring the information, classifying and storing attribute features, behavioral trends, and potential needs separately to form a three-dimensional feature set. For multiple features within the same dimension, they are ranked according to importance and confidence level, with high-value features being retained first. Through this step, the natural language customer profile report is transformed into a structured business feature representation, providing a clear data foundation for subsequent insight analysis and decision support.

[0225] After extracting customer attributes, behavioral trends, and potential needs, business analysis and value interpretation are required to generate customer insight analysis. Insight analysis is not simply a list of features, but a comprehensive business interpretation that involves correlation analysis, value assessment, and opportunity identification of customer characteristics.

[0226] Customer value assessment is the first part of insight analysis. It comprehensively considers a customer's asset size, income level, product holdings, and transaction activity to calculate the customer's current and potential value. Current value is quantified by indicators such as the customer's contributed deposit balance, loan interest, and fee income, while potential value is predicted based on the customer's growth characteristics, such as income growth trends, age group, and career development potential. The value assessment results categorize customers into high-value, medium-value, and low-value tiers, providing a basis for differentiated service strategies.

[0227] Demand forecasting is the second part of insight analysis. Based on customer lifecycle stages, behavioral trends, and potential demand signals, it predicts the financial service needs customers may have in the future. The forecasting process combines rule-based reasoning with machine learning models. The rule engine processes explicit demand signals, such as customers proactively inquiring about a certain type of product, while the machine learning model uncovers implicit demand tendencies from customer characteristics and behavioral patterns. For example, for customers with stable income, children, and children about to start school, it predicts their potential need for education savings.

[0228] Marketing opportunity identification is the third part of insight analysis. Based on customer demand forecasts and product holdings, it identifies specific product recommendation opportunities and marketing entry points. The identification process matches customer needs with product characteristics, selects suitable product portfolios for the customer, and assesses the probability of marketing success. For example, for customers with financial needs and a moderate risk appetite, it identifies opportunities to recommend stable financial products and mixed funds.

[0229] Risk warnings are the fourth part of the insight analysis. For clients with potential risk signals, the analysis covers the type and severity of the risk, along with corresponding recommendations. Risk types include credit risk (e.g., excessively high debt ratio), liquidity risk (e.g., tight cash flow), and market risk (e.g., excessively concentrated investments).

[0230] The insights and analytics are presented in a structured report format, including modules such as customer value rating, demand forecast list, marketing opportunity ranking, and risk warnings. Each module is accompanied by detailed analytical basis and data support. Through this step, customer profile reports are transformed into decision-making recommendations that can directly guide business actions, providing account managers with clear work directions and execution strategies, significantly improving the accuracy of customer service and the conversion rate of marketing campaigns.

[0231] After generating customer insight analysis, it needs to be further transformed into communication scripts that account managers can use directly, i.e., personalized scripts.

[0232] The script generation process begins by analyzing the customer's communication history using an intent recognition model to identify the customer's current communication intent. This intent recognition model is a classification model trained using natural language understanding technology, capable of identifying the customer's true needs and communication objectives from their historical dialogue records, consultation content, and service requests.

[0233] Common communication intentions include product consultation intentions, such as inquiring about the yield of wealth management products or loan application conditions; problem-solving intentions, such as handling complaints or checking account anomalies; business processing intentions, such as applying for a card or purchasing products; and information acquisition intentions, such as policy consultations or market information.

[0234] The intent recognition process analyzes recent customer communication records, extracts keywords and semantic features, and uses a pre-trained intent classifier to determine the dominant intent in the current communication scenario. For example, if a customer has repeatedly inquired about mortgage interest rates and application conditions recently, the intent to inquire about loans is identified. The intent recognition results include intent type, confidence score, and related specific questions, providing contextual information for subsequent strategy decisions.

[0235] Then, through the agent decision-making model, the optimal recommendation strategy is determined by integrating customer insight analysis and current communication intent. The agent decision-making model uses reinforcement learning or decision tree algorithms to learn the optimal decision-making strategy through multiple rounds of interaction, with the goal of maximizing customer satisfaction and business conversion rate.

[0236] The decision-making process considers multiple factors: the customer's need matching degree, i.e. how well the recommended product matches the customer's needs; the customer's acceptance probability is based on historical marketing response rates and current communication intentions; the product's business value, including profit contribution and strategic importance; and the appropriateness of timing, such as whether it is a promotional period or whether the customer has sufficient time.

[0237] The decision-making model outputs recommendation strategies, including recommended product types, recommendation order, communication focus, and communication style.

[0238] Next, the system retrieves product information and sales script templates that match the recommendation strategy from a financial knowledge base. This knowledge base stores detailed information on all financial products, including product name, product type, yield, risk level, investment period, minimum investment amount, target customer group, and other attributes, as well as a library of sales script templates tailored to different scenarios and customer groups.

[0239] The search process retrieves matching products and sales scripts based on product type and customer characteristics in the recommendation strategy. The sales script templates include a complete communication process module, such as opening remarks, needs assessment, product introduction, advantages explanation, objection handling, and closing the deal. Each module has multiple optional versions to suit different customer types and communication styles.

[0240] Finally, based on the personalized information in the customer profile report, the dialogue template was customized. The customization process involved replacing generic placeholders in the template with the customer's specific information.

[0241] The adjustments also include selecting appropriate wording based on customer characteristics: using technical terms and data analysis for highly specialized customers, and using easy-to-understand analogies and case studies for ordinary customers.

[0242] The generated personalized communication scripts are presented in a structured script format, including recommended scripts for each communication stage, key information prompts, and precautions. Account managers can directly refer to them or flexibly adjust them according to actual situations. This step transforms abstract customer insights and recommendation strategies into concrete and actionable communication plans, significantly reducing the workload of account managers, improving communication efficiency and marketing success rates, and achieving a complete closed loop from data analysis to business action.

[0243] The generation of personalized dialogue fully demonstrates the practical value of intelligent decision support systems, providing not only analytical conclusions but also action guidelines.

[0244] Personalization of communication techniques is reflected in multiple aspects: content personalization is based on customized recommended content according to the specific characteristics and needs of the customer; style personalization is based on adjusting the expression according to the customer's communication preferences and comprehension ability; timing personalization is based on selecting appropriate entry points according to the customer's current state and communication intentions; and strategy personalization is based on developing differentiated communication strategies according to the customer's value level and response history.

[0245] The script generation process also takes compliance requirements into account, automatically filtering out statements that may cause controversy or violate regulations, and adding necessary risk warnings and disclaimers.

[0246] The generated scripts are displayed on the account manager's workbench, supporting real-time viewing, online editing, and feedback collection, forming a closed-loop mechanism for continuous optimization. After using the scripts, account managers can annotate their effectiveness and customer reactions. This feedback data flows back into the system to optimize the intent recognition model, decision-making model, and script template library, enabling the system's self-evolution and continuous improvement. In summary, this completes the transformation from data to insight to action, truly bringing the value of the intelligent decision support system to the front line of business, empowering account managers, and enhancing the financial institution's customer service capabilities and market competitiveness.

[0247] This embodiment achieves intelligent transformation from customer profiling to business action through customer insight analysis and personalized script generation mechanisms. Feature extraction structures key information in the profiling report, insight analysis provides in-depth value interpretation and opportunity identification, and script generation provides directly usable communication scripts. The collaborative work of intent recognition, intelligent decision-making, and knowledge retrieval ensures the relevance and effectiveness of the scripts, while personalized adjustments ensure the scripts are fully tailored to customer characteristics and communication scenarios. Compared to traditional manual analysis and experience-based communication, this method provides scientific decision-making suggestions and standardized communication solutions based on data-driven insights, significantly improving the work efficiency and marketing success rate of account managers. Simultaneously, through a continuous feedback optimization mechanism, the system achieves self-evolution, providing strong technical support for the intelligent operation and refined management of financial institutions, effectively enhancing customer experience and business value creation capabilities.

[0248] This application also provides an intelligent decision support system that integrates multi-source financial data and security fences, including: The memory is configured to store instructions; and The processor is configured to retrieve instructions from memory and, when executing instructions, to implement the aforementioned intelligent decision support method that integrates multi-source financial data with security fences.

[0249] This application also provides a machine-readable storage medium storing instructions that cause a machine to execute the aforementioned intelligent decision support method that integrates multi-source financial data and security fences.

[0250] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0251] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0252] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0253] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0254] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0255] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An intelligent decision support method integrating multi-source financial data and security fences, characterized in that, include: In response to receiving a customer profile generation request, the customer profile generation request is parsed, the profile task scenario is determined, and the profile task scenario is converted into a task context vector, wherein the task context vector includes the target customer identifier; Based on the target customer identifier, the original data is retrieved from multiple financial data sources, and the original data is decomposed into multiple feature atoms, wherein each feature atom is an indivisible smallest unit of information. Label metadata for each of the aforementioned feature atoms; A financial feature topology network is constructed based on multiple feature atoms, with the feature atoms as nodes, and connections are established between nodes according to financial knowledge graphs and data relevance. A dynamic confidence score is calculated for each feature atom on the financial feature topology network, and the dynamic confidence score is dynamically adjusted based on the task context vector; Based on the dynamic confidence score, the multiple feature atoms are assigned to different confidence regions of the structured prompt word template to generate structured prompt words. The structured prompt word template includes a high-confidence fact region, a medium-confidence inference region, and a low-confidence reference region. The structured prompts are input into a large language model for inference to generate a customer profile report; Perform a traceability consistency audit on the customer profile report to verify whether the key conclusions in the customer profile report can be traced back to the feature atoms of the high-confidence fact area; Based on customer profile reports that have passed the traceability consistency audit, customer insight analysis and personalized communication scripts are generated and displayed through the account manager's workbench.

2. The method according to claim 1, characterized in that, Based on the target customer identifier, original data is retrieved from multiple financial data sources, including: Based on the target customer identifier, internal business data is extracted from a pre-set internal tagging system. The internal business data includes wealth data, credit data, and payment data. Obtain anonymized external data from third-party data sources through a privacy-preserving computing interface; The internal business data and the external data are fused to retrieve the original data.

3. The method according to claim 1, characterized in that, A financial feature topology network is constructed based on multiple of the aforementioned feature atoms, including: Multiple of the aforementioned feature atoms are used as network nodes; Extract product relationships and behavioral causal relationships from a pre-defined financial knowledge graph; Determine the temporal proximity and logical consistency among the characteristic atoms; Based on the product association, the behavioral causal relationship, and the temporal proximity and logical consistency between the feature atoms, edge connections are established between the network nodes; The financial feature topology network is generated based on the network nodes and the edge connections.

4. The method according to claim 1, characterized in that, Calculating the dynamic confidence score for each feature atom on the financial feature topology network includes: A preset feature vector centrality algorithm is used to calculate the static centrality score of each feature atom on the financial feature topology network. The static centrality score reflects the objective importance of the feature atom in the financial feature topology network. The task context vector is applied to the financial feature topology network to adjust the weights of the edge connections related to the portrait task scenario, resulting in the adjusted financial feature topology network. The centrality score of each feature atom is recalculated on the adjusted financial feature topology network to obtain the dynamic confidence score of each feature atom.

5. The method according to claim 4, characterized in that, Applying the task context vector to the financial feature topology network to adjust the weights of edge connections related to the profiling task scenario includes: Identify edge connections in the financial feature topology network that are semantically related to the portrait task scenario; When the profiling task scenario is marketing response evaluation, the weights of edge connections related to consumer behavior and interest preferences are dynamically increased. When the portrait task scenario is credit risk assessment, the weight of edge connections related to debt information, cash flow, and repayment history is dynamically increased.

6. The method according to claim 1, characterized in that, Based on the dynamic confidence score, the multiple feature atoms are assigned to different confidence regions of the structured cue word template, including: Feature atoms with dynamic confidence scores greater than the high confidence threshold are assigned to the high confidence fact area, and a first instruction identifier is added to the high confidence fact area. The first instruction identifier is used to instruct the large language model to use the feature atoms in the high confidence fact area as core factual evidence. Feature atoms whose dynamic confidence scores fall between the low confidence threshold and the high confidence threshold are assigned to the medium confidence inference region, and a second instruction identifier is added to the medium confidence inference region. The second instruction identifier is used to instruct the large language model to analyze the feature atoms in the medium confidence inference region as important clues. Feature atoms with dynamic confidence scores less than the low confidence threshold are assigned to the low confidence reference region, and a third instruction identifier is added to the low confidence reference region. The third instruction identifier instructs the large language model to use the feature atoms in the low confidence reference region as background reference information.

7. The method according to claim 1, characterized in that, Perform a traceability consistency audit on the customer profile report, including: Extract key conclusion statements from the customer profile report; Perform semantic analysis on each of the key conclusion statements to determine the core argument of each key conclusion statement; Retrieve supporting evidence related to the core argument from the characteristic atoms of the high-confidence fact region; If the supporting evidence is found, the key conclusion statement is marked as a high-confidence conclusion. If the supporting evidence is not found, the key conclusion statement is marked as a low-confidence inference, and the key conclusion statement is removed from the customer profile report and / or an uncertainty label is added.

8. The method according to claim 1, characterized in that, Based on customer profile reports that have passed the traceability consistency audit, customer insight analysis and personalized communication scripts are generated, including: Extract customer attribute characteristics, behavioral trends, and potential needs from customer profile reports that have passed the traceability consistency audit; Based on the customer attribute characteristics, behavioral trends, and potential needs, customer insight analysis is generated. The system receives customer insight analysis and combines it with an intent recognition model, an intelligent agent decision-making model, and a financial knowledge base to generate personalized dialogue.

9. An intelligent decision support system integrating multi-source financial data and security fences, characterized in that, include: The memory is configured to store instructions; as well as A processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the intelligent decision support method for integrating multi-source financial data and security fences according to any one of claims 1 to 8.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute the intelligent decision support method for integrating multi-source financial data and security fences according to any one of claims 1 to 8.

Images