Banking intelligent decision system based on bi and ai double driving

Abstract

The application relates to the field of intelligent decision-making, and discloses a bank intelligent decision-making system based on BI and AI double driving, which comprises a data production layer, a data transmission layer, a data consumption layer, and a business action layer. The data production layer is used for collecting multi-source bank business data and generating data dashboards corresponding to business indexes. The data transmission layer is used for distributing the data dashboards to user terminals. The data consumption layer comprises a multi-terminal adaptation module and an AI data question and answer module. The multi-terminal adaptation module is used for displaying the data dashboards on different types of user terminals. The AI data question and answer module is used for receiving natural language queries initiated by users for the business indexes, analyzing query intentions through multiple artificial intelligences, and generating visual reports. The business action layer is used for monitoring the execution state of to-be-completed tasks, and when the execution state is a completed state, obtaining a treatment conclusion and feeding back the treatment conclusion to the data production layer. The scheme effectively improves the real-time performance of bank data integration and the business landing efficiency.

Description

Technical Field

[0001] This manual relates to the field of intelligent decision-making, and in particular to intelligent decision-making systems for the banking industry driven by both BI and AI. Background Technology

[0002] With the deepening of fintech and digital transformation, the banking industry is accelerating its transformation towards digitalization and intelligence. Business systems are becoming increasingly diversified and complex, and data volume is growing exponentially, placing core demands on bank decision-making to be "real-time, accurate, and executable." However, in the current field of bank decision-making, data application faces the following challenges.

[0003] Business data is scattered across dozens of heterogeneous systems, including core banking, credit, risk control, and trading systems, forming data silos. Cross-system integration, cleaning, and single collection can take 2-3 days, failing to support real-time decision-making needs. Traditional data analysis heavily relies on technical personnel writing SQL statements and creating reports, with an average business demand response cycle of 5-7 days, making it difficult for non-technical personnel to independently complete data analysis. Multi-terminal collaboration and proactive early warning mechanisms are lacking. Existing systems only support viewing fixed reports on PCs, making it impossible for senior executives to keep abreast of business dynamics when they are away. Anomalies in indicators require manual inspection, resulting in significant delays in risk detection. The data value conversion rate is low, with data mostly used for "post-event statistics." There is a lack of effective linkage mechanisms between data discovery and business actions, leading to a disconnect between decision-making and action. Early warnings are limited to message pushes and lack linkage with pending tasks and workflows, resulting in a dilemma of "early warnings but no one to manage them, data but no action."

[0004] Therefore, there is an urgent need for a banking intelligent decision-making system driven by both BI and AI to achieve structured integration of multi-source business data and accurate generation of business indicators, thereby achieving a closed loop from data production to business action and solving the shortcomings of traditional technologies in data utilization, intelligent interaction, and decision execution. Summary of the Invention

[0005] In view of this, the present invention aims to propose a banking intelligent decision-making system driven by both BI and AI, in order to solve the problems of long data response cycles, lack of intelligent interaction capabilities, and gaps in decision-making and execution caused by the separation of data production, consumption and action links in the existing banking data system and the lack of deep integration of business intelligence and artificial intelligence technologies, so as to realize closed-loop intelligent decision-making throughout the entire process from data collection to business action.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A banking intelligent decision-making system driven by both BI and AI includes: The data production layer is used to collect multi-source banking business data, classify and label the multi-source banking business data through a business intelligence agent to generate business indicators, and generate data dashboards corresponding to the business indicators. The data transmission layer is used to distribute the data dashboard to user terminals; The data consumption layer includes a multi-terminal adaptation module and an AI data question-and-answer module. The multi-terminal adaptation module is used to display the data dashboard on different types of user terminals. The AI ​​data question-and-answer module is used to receive natural language queries initiated by users for the business indicators, parse the query intent through multiple artificial intelligence agents, obtain the corresponding business indicators from the data production layer, and generate a visual report. The business action layer is used to associate early warning information with pending tasks, create and assign pending tasks according to preset business rules, monitor the execution status of pending tasks, and obtain a handling conclusion when the execution status is completed. The handling conclusion is then fed back to the data production layer and stored in conjunction with the business indicators that triggered the generation of the early warning information.

[0007] The beneficial effects of this solution are as follows: By constructing a closed-loop architecture encompassing a data production layer, data transmission layer, data consumption layer, and business action layer, this system achieves deep integration of business intelligence and artificial intelligence, solving the problems of long data response cycles, lack of intelligent interaction capabilities, and gaps in decision-making and execution in existing bank data systems. The data production layer classifies and tags multi-source data to generate business indicators and visualizes them as data dashboards. The data consumption layer, through multi-terminal adaptation and an AI data question-and-answer module, supports natural language queries and generates visual reports. The business action layer associates early warning information with pending tasks, monitors execution status, and feeds back the handling conclusions to the data production layer, binding them with business indicators for storage. This system realizes a closed-loop intelligent decision-making process from data collection to business action, effectively improving data response efficiency, intelligent interaction level, and decision-making execution effectiveness.

[0008] Furthermore, the data production layer includes: Data storage unit, used to store multi-source banking business data; The data acquisition unit adopts a distributed architecture to collect multi-source banking business data from multiple source systems of the bank, and stores the collected multi-source banking business data in the data storage unit; The data processing unit classifies and tags the multi-source banking business data through a business intelligence agent to generate business indicators; The dashboard generation unit is used to visualize the business indicators into a data dashboard according to personalized needs. The dashboard generation unit includes a visualization component library and a drag-and-drop interactive interface. The drag-and-drop interactive interface is used to receive drag-and-drop operation instructions from users on the components in the visualization component library, and dynamically adjust the layout and display content of the data dashboard based on the drag-and-drop operation instructions.

[0009] Beneficial effects: By setting up data storage units, data acquisition units, data processing units, and dashboard generation units, the specific composition of the data production layer is clarified. The data acquisition unit adopts a distributed architecture to collect and store data from multiple source systems. The data processing unit uses a business intelligence agent to classify and label data to generate business indicators. The dashboard generation unit realizes personalized generation and dynamic adjustment of data dashboards through a visual component library and drag-and-drop interactive interface, which improves the efficiency of data collection, the accuracy of indicator generation, and the flexibility of dashboard configuration.

[0010] Furthermore, the data transmission layer includes a real-time push mode and an on-demand pull mode; The real-time push mode is used to push early warning information or visual reports to the user terminal when abnormal business indicators are detected. The on-demand retrieval mode is used to respond to proactive data query requests initiated by users through the user terminal. Based on the proactive data query request, the corresponding business indicators or data dashboards are obtained from the data production layer, and the results are returned to the user terminal that initiated the request for display.

[0011] Beneficial effects: By setting up real-time push mode and on-demand retrieval mode, the data transmission layer can flexibly distribute data according to business needs. In real-time push mode, when business indicators change, early warning information or visual reports are proactively pushed to user terminals to ensure that key information is delivered in a timely manner. In on-demand retrieval mode, users actively query and return the corresponding business indicators or data dashboards to meet users' needs for independent data acquisition, thereby improving the timeliness of data transmission and the convenience of interaction.

[0012] Furthermore, the AI ​​data question-answering module includes multiple AI agents, which adopt a cascaded processing architecture, including: An intent recognition intelligent agent is used to parse natural language queries initiated by users for the aforementioned business metrics and identify their query intent, time range, and target metrics. A knowledge retrieval agent is used to retrieve indicator definitions and calculation rules related to the query intent from the enterprise knowledge base; SQL generation agent, used to automatically generate data query statements based on the outputs of the intent recognition agent and the knowledge retrieval agent; A data query agent is used to execute the data query statement and obtain the corresponding business indicators from the data production layer. The results interpretation agent is used to perform trend analysis and business interpretation on the acquired business indicators. The report generation agent integrates the output of the result interpretation agent to generate a visual report containing key conclusions, detailed data, and action recommendations.

[0013] Beneficial effects: By setting up multiple intelligent agents such as intent recognition, knowledge retrieval, SQL generation, data query, result interpretation, and report generation, and adopting a cascaded processing architecture, the AI ​​data question answering module can parse users' natural language queries layer by layer, accurately identify query intent, and automatically generate data query statements. After obtaining business indicators from the data production layer, it performs trend analysis and business interpretation, and finally generates a visual report containing key conclusions and action suggestions, which significantly improves the processing accuracy and response efficiency of complex queries.

[0014] Furthermore, the business action layer includes: The early warning and monitoring module is used to continuously monitor business indicators and generate early warning information when the business indicator value triggers a preset threshold. The task linkage module is used to receive the warning information, create task assignments according to the warning type and preset business rules, and assign the task assignments to the corresponding personnel. The task monitoring module is used to monitor the execution status of the pending tasks and to provide upgrade reminders for pending tasks whose execution status has expired and has not been updated. The result feedback module is used to obtain the handling conclusion when the execution status of the pending task is completed, and to feed the handling conclusion back to the data production layer, and bind and store it with the business indicator that triggered the generation of the early warning information.

[0015] Beneficial effects: By setting up an early warning monitoring module, a to-do linkage module, a task monitoring module, and a result feedback module, the business action layer realizes a closed loop of the entire process from early warning triggering to task processing and result feedback. The early warning monitoring module generates early warning information when business indicators trigger thresholds. The to-do linkage module automatically creates to-do tasks according to the early warning type and assigns them to the corresponding personnel. The task monitoring module tracks the execution status in real time and reminds overdue tasks. The result feedback module binds and stores the handling conclusions with the business indicators that triggered the early warning, ensuring that the early warning information is effectively processed and forms a traceable handling record.

[0016] Furthermore, it also includes a knowledge base self-evolution module, which includes a cold start phase, a continuous learning phase, and an error correction phase; When the knowledge base self-evolution module is in the cold start phase, historical data query templates and business indicator definition files are imported to build the basic knowledge base. When the knowledge base self-evolution module is in the continuous learning phase, it extracts new business rules from the dialogue between the user and the AI ​​data question-and-answer module and updates the knowledge base. When the knowledge base self-evolution module is in the error correction phase, it receives user feedback on the visualization report and updates the knowledge base according to the correction instructions.

[0017] Beneficial effects: By setting up a self-evolving knowledge base module and dividing it into three stages—cold start, continuous learning, and error correction—the system can quickly build a basic knowledge base and continuously optimize it during use. In the cold start stage, historical data query templates and business indicator definition files are imported to lay the knowledge foundation. In the continuous learning stage, new business rules are extracted from user dialogues and the knowledge base is updated. In the error correction stage, the knowledge base content is corrected based on user feedback on the visualization report. This achieves dynamic updates and self-improvement of the knowledge base, enhancing the accuracy and adaptability of AI question answering.

[0018] Furthermore, the multi-terminal adaptation module includes: Terminal type identification unit, used to obtain the type of terminal device currently logged in by the user; The content adaptive rendering unit is communicatively connected to the terminal type identification unit and is configured to render a complete data dashboard that supports multi-dimensional drill-down when the terminal device type is identified as a web page; and extract business indicators from the data dashboard and perform lightweight rendering according to mobile interaction specifications when the terminal device type is identified as a mobile device. The context synchronization unit is used to synchronize the current query context, filter conditions, and alert information between different terminals when a user switches terminal devices.

[0019] Beneficial effects: By setting up a terminal type recognition unit, a content adaptive rendering unit, and a context synchronization unit, the multi-terminal adaptation module can automatically adjust the rendering method of the data dashboard according to the type of terminal device the user logs in with. The web terminal renders a complete multi-dimensional drill-down dashboard, while the mobile terminal extracts business indicators for lightweight rendering. When the user switches terminals, the module can simultaneously query the context, filter conditions, and warning information, thus achieving a consistent experience and seamless connection across terminals.

[0020] Furthermore, the content adaptive rendering unit includes: The rendering strategy experiment pool is used to store multiple candidate rendering strategies preset for different terminal types and user query scenarios; A rendering decision scorer is used to obtain user operation data of different candidate rendering strategies on corresponding terminal types in historical data, and to calculate the experience score of each candidate rendering strategy based on the user operation data. The rendering strategy selection subunit is connected to the terminal type identification unit and the rendering decision scorer. It is used to select the candidate rendering strategy with the highest experience score from the rendering strategy experiment pool as the target rendering strategy according to the current terminal device type. A dynamic rendering executor is used to render the data dashboard according to the target rendering strategy.

[0021] Beneficial effects: By setting up a rendering strategy experiment pool, a rendering decision scorer, a rendering strategy optimization sub-unit, and a dynamic rendering executor, the content adaptive rendering unit can quantitatively score different candidate rendering strategies based on historical user operation data, and select the target rendering strategy to execute rendering according to the current terminal type. This achieves data-driven optimization of rendering strategies and improves the user's interactive experience on different terminals.

[0022] Furthermore, the rendering decision scorer calculates the experience score using the following logic: ; In the formula, Indicates the first Experience scores for candidate rendering strategies on terminal type j; This represents the click-through rate, which is the ratio of valid interactions to total interactions. This represents the page dwell time index, which is a normalized value of the actual dwell time and the baseline time. This represents the task completion rate, which is the ratio of the number of times the target operation was completed to the total number of visits. This represents the early exit rate, which is the ratio of the number of visits with a dwell time below a preset threshold to the total number of visits. , These are the preset weighting coefficients.

[0023] Beneficial effects: By setting a specific calculation formula for the experience score, the user operation data from multiple dimensions, such as interaction click pass rate, page dwell time index, task completion rate, and early exit rate, are weighted and calculated, providing a clear quantitative basis for the selection of rendering strategies, ensuring the objectivity and repeatability of the scoring process, and improving the scientificity and accuracy of rendering strategy selection. Attached Figure Description

[0024] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary module diagram of a banking intelligent decision-making system driven by both BI and AI. Detailed Implementation

[0025] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0026] As indicated in this specification and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0027] Flowcharts are used in this specification to illustrate the operations performed by the system according to embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0028] The following detailed explanation illustrates the specific implementation methods: Example 1: like Figure 1 As shown, the banking intelligent decision-making system, driven by both BI and AI, includes: The data production layer is used to collect multi-source banking business data, classify and label the multi-source banking business data through a business intelligence agent to generate business indicators, and generate data dashboards corresponding to the business indicators.

[0029] Furthermore, the data production layer includes: Data storage unit, used to store multi-source banking business data.

[0030] In this embodiment, the data storage unit is constructed using a distributed data warehouse. Specifically, the underlying storage is implemented based on the Hadoop Distributed File System, and an offline data warehouse is built using Hive to support batch processing analysis. At the same time, ClickHouse is deployed to build a real-time data warehouse to support high-concurrency, low-latency ad-hoc query requirements.

[0031] Multi-source banking business data refers to the collection of various raw data generated by banks during their operations and stored in different business systems.

[0032] For example, transaction data from the banking system, credit and repayment records from the credit system, anti-fraud lists and blacklists from the risk control system, mobile banking and online banking behavior logs from the channel system, and credit data and business information from external procurement.

[0033] The data acquisition unit adopts a distributed architecture to collect multi-source banking business data from multiple source systems of the bank, and stores the collected multi-source banking business data in the data storage unit.

[0034] In this embodiment, the system adopts a distributed architecture and uses data acquisition components such as Kafka and Flume to collect multi-source banking business data from multiple source systems in real time or near real time, ensuring the reliability and throughput of data transmission.

[0035] Multiple source systems refer to various production systems within a bank that operate independently and generate their own business data.

[0036] For example, the core banking system is responsible for basic transactions such as account opening, transfers, deposits and withdrawals; the credit system is responsible for loan application, approval, disbursement and repayment processes; the risk control system is responsible for anti-fraud, blacklist management and abnormal transaction monitoring; channel systems, such as online banking, mobile banking, ATMs and counter systems; and external data source interfaces, such as credit reporting agencies and government data platforms.

[0037] Distributed architecture refers to a cluster system composed of multiple servers that achieves efficient collection, storage, and computation of massive amounts of data through parallel processing and task collaboration. Specifically, it can use Hadoop HDFS for distributed storage, Spark or Flink for distributed computing and data processing, and Kafka for distributed message queues to support high-throughput data access. This architecture has horizontal scalability, allowing nodes to be dynamically added according to the data scale, ensuring the system's stability and processing performance under large data volumes.

[0038] In this embodiment, after the data acquisition unit collects multi-source banking business data, the data will be cleaned, transformed and loaded through an ETL task. Specifically, this includes: cleaning missing values ​​and outliers, converting the data formats of different source systems into a unified standard format, mapping fields and transforming business logic according to a predefined data model, and finally loading the processed data into the data storage unit.

[0039] The predefined data model refers to the data warehouse model built using dimensional modeling methods. This model consists of fact tables and dimension tables. Fact tables store measurement events for business processes, including foreign keys and numeric measurement fields associated with dimension tables, such as loan amount and number of transactions. Dimension tables provide the context for business processes, containing descriptive attribute fields such as time, customer, and product. The fact tables and dimension tables are organized using a star schema, with the fact table at the center and multiple dimension tables directly linked through foreign keys, forming a data structure that supports multi-dimensional analysis.

[0040] The data processing unit uses a business intelligence agent to classify and label multi-source banking business data, generating business indicators.

[0041] Business metrics are calculable values ​​used to quantitatively assess the status, performance, and risk level of banking operations. They are abstract expressions of business operations after processing the raw data.

[0042] For example, non-performing loan ratio, average daily deposit balance, customer conversion rate, and proportion of risky assets.

[0043] Business metrics are calculable values ​​used to quantitatively assess the status, performance, and risk level of a banking business. Examples include a non-performing loan ratio of 3.5%, an average daily deposit balance of 120 million yuan, and a customer conversion rate of 15% this month.

[0044] In this embodiment, the business intelligence agent classifies multi-source banking business data according to business themes, dividing them into thematic domains such as customer domain, product domain, and risk domain. Based on the classification, business tags are added to the data, such as "new customer", "large transaction" and "non-performing". Based on the classification results and tag information, quantifiable business indicators are automatically generated through preset calculation rules.

[0045] The preset calculation rules can include summation, counting, compound operations, conditional aggregation, trend calculation, etc.

[0046] The dashboard generation unit is used to visualize business metrics into data dashboards according to personalized needs. The dashboard generation unit includes a visualization component library and a drag-and-drop interactive interface. The drag-and-drop interactive interface is used to receive drag-and-drop operation instructions from users on the components in the visualization component library, and dynamically adjust the layout and display content of the data dashboard based on the drag-and-drop operation instructions.

[0047] A data dashboard is an interactive interface that displays business metrics in a centralized manner using visual formats such as charts, tables, and dashboards.

[0048] In this embodiment, the data dashboard may include multiple visualization components, such as line charts to show trend changes, bar charts to compare data in different dimensions, tables to show detailed data, and dashboards to show the progress of key indicators.

[0049] In this embodiment, the user selects the desired chart type, indicator field, and dimension field from the visualization component library through a drag-and-drop interactive interface. The system generates the corresponding visualization component in real time according to the user's operation instructions and places it at the specified position on the canvas. The user can continue to drag and adjust the position, size, and layout of the components, and the system responds in real time and updates the dashboard display effect. After completing the layout, the user can save the current configuration to generate a personalized data dashboard. The system stores the dashboard configuration information in the database and loads and renders it on demand during subsequent accesses.

[0050] The dashboard configuration information may include component type, bound business metrics, layout parameters, etc.

[0051] The data transmission layer is used to distribute data dashboards to user terminals.

[0052] User terminals are the front-end devices and applications running on the bank's intelligent decision-making system that allow users to access and interact with it. User terminals include web-based and mobile terminals.

[0053] Web-based applications refer to web applications accessed via a browser on a personal computer, supporting complex data dashboard analysis, multi-dimensional data drill-down, and drag-and-drop self-service analysis operations.

[0054] Mobile devices refer to applications that run on mobile devices such as smartphones and tablets, including lightweight applications of collaborative office software such as DingTalk and WeChat Work, as well as standalone mobile apps.

[0055] Furthermore, the data transmission layer includes a real-time push mode and an on-demand pull mode.

[0056] The real-time push mode is used to push early warning information or visual reports to user terminals when abnormal business indicators are detected.

[0057] An abnormal business indicator means that the current value of the business indicator deviates from the preset normal range.

[0058] In this embodiment, the anomaly can be derived by a real-time computing engine, such as Flink, which performs streaming computation on the incoming business indicator data and dynamically matches it with a preset rule base.

[0059] Early warning information refers to a structured notification message automatically generated by the system when business metrics change, containing details of the anomaly and handling instructions. In this embodiment, the warning information may include the name of the abnormal indicator, the current indicator value, the preset threshold, the time of the abnormality, a description of the abnormal trend, and suggested handling measures. The warning information is packaged in a standard format for easy distribution to different terminals and subsequent linkage with pending tasks.

[0060] In this embodiment, the early warning monitoring module continuously monitors business metrics and uses a complex event processing engine to calculate the matching relationship between each metric value and a preset threshold in real time. When an anomaly condition is met, the engine triggers an early warning event, calls the early warning information generation service, queries the knowledge base for the relevant processing rules and responsible person information for that metric, fills in the various fields in the early warning information template, and generates a complete early warning information object. The generated early warning information is distributed to user terminals in real-time push mode and persistently stored in an early warning record table for subsequent to-do task creation and historical tracing.

[0061] The on-demand retrieval mode is used to respond to proactive data query requests initiated by users through user terminals. Based on the proactive data query requests, the corresponding business indicators or data dashboards are retrieved from the data production layer, and the retrieved results are returned to the user terminal that initiated the request for display.

[0062] In this embodiment, by setting up a real-time push mode and an on-demand retrieval mode, the problems of existing bank data systems that only support viewing fixed reports, cannot proactively push abnormal information, and users cannot independently obtain data anytime and anywhere are solved. This realizes a flexible combination of proactive delivery of key information and on-demand query by users, improving the timeliness of data transmission and the convenience of interaction.

[0063] The data consumption layer includes a multi-terminal adaptation module and an AI data question-and-answer module. The multi-terminal adaptation module is used to display data dashboards on different types of user terminals. The AI ​​data question-and-answer module is used to receive natural language queries initiated by users for business indicators, analyze the query intent through multiple artificial intelligence agents, obtain the corresponding business indicators from the data production layer, and generate visual reports.

[0064] Furthermore, the multi-terminal adaptation module includes: The terminal type identification unit is used to obtain the type of terminal device that the user is currently logged into.

[0065] The content adaptive rendering unit is connected to the terminal type recognition unit and is configured to render a complete data dashboard that supports multi-dimensional drill-down when the terminal device type is identified as a web page; and to extract business indicators from the data dashboard and perform lightweight rendering according to the mobile interaction specifications when the terminal device type is identified as a mobile device.

[0066] Furthermore, the content-adaptive rendering unit includes: The rendering strategy experiment pool is used to store multiple candidate rendering strategies preset for different terminal types and user query scenarios.

[0067] User query scenarios refer to the business context and intent type when a user initiates a data query. This reflects the different data presentation needs of users at different stages of their work.

[0068] For example, executives need to quickly browse the summary dashboard of core operating indicators during daily morning meetings; risk control specialists need to view detailed trend charts and data of risk indicators when investigating anomalies.

[0069] Pre-defined multiple candidate rendering strategies refer to multiple different presentation schemes pre-designed for the same terminal type and query scenario.

[0070] For example, for mobile core metric query scenarios, candidate rendering strategies can include card-style layout, dashboard-style layout, or hybrid layout.

[0071] Among them, the card layout arranges business indicators vertically in the form of cards, with each card displaying the indicator name, current value, month-on-month change, and trend icon, and supports left and right swiping to switch indicator categories; the dashboard layout displays key indicator dashboards at the top, trend line charts in the middle, and detailed indicator list at the bottom, and supports drill-down; the hybrid layout uses a collapsible group layout, which displays the indicator category title and summary value by default, and displays detailed data and charts of all indicators under that category when clicked to expand.

[0072] The rendering decision scorer is used to obtain user operation data of different candidate rendering strategies on corresponding terminal types from historical data, and to calculate the experience score of each candidate rendering strategy based on the user operation data.

[0073] The experience score is an indicator used to quantitatively evaluate the user experience of different rendering strategies under specific terminal types and user query scenarios; it reflects the user's acceptance of the data dashboard presentation method and the efficiency of interaction.

[0074] In this embodiment, the experience score is calculated based on user operation data. The higher the score, the more the rendering strategy meets the user's expectations and operating habits in the current scenario.

[0075] User interaction data includes various behavioral records generated during user interactions with the data dashboard, such as click count, page dwell time, operation path for completing specific tasks, scroll depth, and exit timing. Experience scoring transforms subjective user experience into objective, comparable values ​​by weighting these multi-dimensional user interaction data, providing a quantitative basis for optimizing rendering strategies.

[0076] Furthermore, when the rendering decision rater calculates the experience score, its calculation logic is as follows: ; In the formula, Indicates the first Experience scores for candidate rendering strategies on terminal type j; This represents the click-through rate, which is the ratio of valid interactions to total interactions. This represents the page dwell time index, which is a normalized value of the actual dwell time and the baseline time. This represents the task completion rate, which is the ratio of the number of times the target operation was completed to the total number of visits. This represents the early exit rate, which is the ratio of the number of visits with a dwell time below a preset threshold to the total number of visits. , These are the preset weighting coefficients.

[0077] In this embodiment, the formula calculates the experience score by quantifying user interaction behavior. Among these metrics is the interaction click-through rate. Page dwell time index reflects the smoothness of user interaction with the data dashboard. Measuring user engagement with content and task completion rate Assess the effectiveness of users in achieving expected goals, including early exit rate. This serves as a negative indicator of user churn. The four dimensions of the indicator are multiplied by their respective weighting coefficients and then summed to obtain a quantitative experience score. A higher score indicates a better user experience for the corresponding terminal type.

[0078] The rendering strategy selection subunit connects the terminal type identification unit and the rendering decision scorer. It is used to select the candidate rendering strategy with the highest experience score from the rendering strategy experiment pool as the target rendering strategy based on the current terminal device type.

[0079] In this embodiment, after receiving the current terminal device type output by the terminal type identification unit, the rendering strategy optimization subunit selects all candidate rendering strategies suitable for the terminal type from the rendering strategy experiment pool; it calls the rendering decision scorer to obtain the historical experience scores of these candidate strategies on the corresponding terminal type, and selects the candidate strategy with the highest experience score as the target rendering strategy by sorting and comparing; if there are multiple candidate strategies with the same score or lack historical score data, the preset default priority is used as the target rendering strategy to ensure that the user can obtain the current optimal rendering scheme every time they visit.

[0080] The dynamic rendering executor is used to render the data dashboard according to the target rendering strategy.

[0081] In this embodiment, the dynamic rendering executor parses the layout structure, component type, and style parameters defined in the target rendering strategy; obtains the business indicator values ​​that need to be displayed from the data production layer and binds them to the corresponding visualization components; calls the underlying visualization rendering engine to generate specific chart elements and positions and draws each component on the canvas according to the layout rules specified by the strategy; and outputs the rendered overall dashboard to the user terminal for display.

[0082] Among them, the underlying visualization rendering engines are such as ECharts and D3.js.

[0083] The context synchronization unit is used to synchronize the current query context, filter conditions, and alert information between different terminals when a user switches terminal devices.

[0084] In this embodiment, when a user logs out of the current terminal, the context synchronization unit automatically captures the query context, filter conditions, and unread alert status in the current session, encapsulates this context data into a structured object, and stores it in the user session cache on the server. When the user logs in again on another terminal, the context synchronization unit reads the user's most recent context data from the cache and pushes it to the data consumption module of the new terminal, so that the data dashboard automatically restores to the interface state of the user's last operation, including the same query results, filter settings, and alert information to be processed, achieving a seamless cross-terminal experience.

[0085] The query context in the current session can include the currently viewed business metrics, time range, etc.; the filtering conditions can include organizational dimension, product dimension filtering conditions, etc.

[0086] In this embodiment, terminal identification and adaptive rendering are used to adapt to the needs of banks' office operations across all scenarios, breaking through the limitations of traditional systems with only one terminal. The context synchronization unit avoids repetitive operations across terminals, improving efficiency. At the same time, the optimal rendering strategy is dynamically selected based on historical data, which fits the characteristics of the terminal and user habits, significantly improving the ease of operation for non-technical personnel and the data consumption experience.

[0087] Furthermore, the AI ​​data question-answering module includes multiple AI agents, which employ a cascaded processing architecture, including: An intent recognition agent is used to parse natural language queries initiated by users for the aforementioned business metrics, and to identify their query intent, time range, and target metrics.

[0088] In this embodiment, a pre-trained language model based on the Transformer architecture is used and trained using intent detection technology for a task-oriented dialogue system. During training, supervised fine-tuning of SFT combined with GRPO reinforcement learning is employed to optimize the model on labeled banking domain query corpora, enabling it to accurately identify specific business intents such as loan inquiries, deposit analysis, and risk monitoring.

[0089] Among them, the pre-trained language models based on the Transformer architecture can be large models such as BERT, GPT series, or Tongyi Qianwen.

[0090] A knowledge retrieval agent is used to retrieve indicator definitions and calculation rules related to the query intent from the enterprise knowledge base.

[0091] In this embodiment, the knowledge retrieval agent is built on the Retrieval Enhanced Generation (RAG) architecture. It stores the indicator definitions, calculation rules and business terms in the enterprise knowledge base in a vectorized manner. When a query intent is received, the agent retrieves the most relevant knowledge content through vector similarity retrieval.

[0092] SQL generation agent is used to automatically generate data query statements based on the outputs of the intent recognition agent and the knowledge retrieval agent.

[0093] In this embodiment, the SQL generation agent adopts a multi-agent Text-to-SQL framework, which decomposes the conversion from natural language to SQL into multiple steps such as pattern linking, sub-problem identification, and query plan generation. It also combines a thought chain reasoning and error classification-guided error correction mechanism to improve the accuracy of SQL generation in complex query scenarios.

[0094] A data query agent is used to execute the data query statement and obtain the corresponding business indicators from the data production layer.

[0095] In this embodiment, the data query agent is responsible for executing the generated SQL statements and interacting with the underlying data warehouse. Its technical implementation is based on a mature database connection and query execution framework. For example, it connects to a Hive or ClickHouse data warehouse through standard database interfaces such as JDBC and ODBC, executes SQL query statements, and retrieves the returned result set.

[0096] The results interpretation agent is used to perform trend analysis and business interpretation on the acquired business indicators.

[0097] The report generation agent integrates the output of the result interpretation agent to generate a visual report containing key conclusions, detailed data, and action recommendations.

[0098] In this embodiment, the result interpretation agent and the report generation agent can be implemented based on existing multimodal report generation frameworks. For example, by adopting a code-driven variable memory (CAVM) architecture and an iterative visual enhancement mechanism, it is possible to perform trend analysis and business interpretation on query results, and generate visual reports that include text and charts.

[0099] In this embodiment, by setting up multiple intelligent agents such as intent recognition, knowledge retrieval, SQL generation, data query, result interpretation, and report generation, and adopting a cascaded processing architecture, the AI ​​data question answering module can parse the user's natural language query layer by layer, accurately identify the query intent and automatically generate data query statements. After obtaining business indicators from the data production layer, it performs trend analysis and business interpretation, and finally generates a visual report containing key conclusions and action suggestions. This solves the problems of traditional data analysis relying on technical personnel to write SQL statements, long response cycles, and difficulty for non-technical personnel to perform queries independently, and significantly improves the processing accuracy and response efficiency of complex queries.

[0100] The business action layer is used to associate early warning information with pending tasks, create and assign pending tasks according to preset business rules, monitor the execution status of pending tasks, and obtain the handling conclusion when the execution status is completed. The handling conclusion is then fed back to the data production layer and stored in conjunction with the business indicators that triggered the generation of early warning information.

[0101] Furthermore, the business action layer includes: The early warning and monitoring module is used to continuously monitor business indicators and generate early warning information when the business indicator value triggers a preset threshold.

[0102] The task linkage module is used to receive early warning information, create tasks based on the warning type and preset business rules, and assign the tasks to the corresponding personnel.

[0103] The task monitoring module is used to monitor the execution status of pending tasks and provide upgrade reminders for pending tasks whose execution status has expired and has not been updated.

[0104] The result feedback module is used to obtain the handling conclusion when the execution status of the pending task is completed, and to feed the handling conclusion back to the data production layer, which is then bound and stored with the business indicators that trigger the generation of early warning information.

[0105] The handling conclusion is a record of the handling results filled in by the person in charge after handling the pending tasks. It includes the root cause analysis of the problem, the handling measures taken, the evaluation of the handling effect, and suggestions for subsequent improvement.

[0106] In this embodiment, the handling conclusion is stored in the form of structured data, including at least fields such as the person handling the handling, the handling time, the handling method, and the handling result, so as to fully trace the entire process of the early warning event from its occurrence to its closure.

[0107] In this embodiment, when the person in charge logs into a collaborative office system (such as OA or DingTalk) and opens the task to be done, the system displays the details of the business indicators associated with the warning information and a preset form for filling in the handling conclusion. The person in charge fills in the handling conclusion in the form and submits it. The result feedback module receives the submitted data, performs format verification and integrity checks on the filled content, saves the handling conclusion that passes the verification to the database, and updates it in association with the execution status of the task to be done.

[0108] In this embodiment, after receiving the handling conclusion, the result feedback module extracts the unique identifier (such as indicator ID, statistical time point, dimension combination) of the business indicator that triggered the generation of the early warning information from the context of the task to be done; using the business indicator identifier as the association key, the handling conclusion is written as a new data entry into the business indicator history table of the data production layer, or a handling conclusion field is added to the metadata record of the business indicator to store the handling result of the most recent early warning, thereby achieving binding.

[0109] Furthermore, this system also includes a knowledge base self-evolution module, which includes a cold start phase, a continuous learning phase, and an error correction phase.

[0110] When the knowledge base self-evolution module is in the cold start phase, import historical data query templates and business indicator definition files to build the basic knowledge base.

[0111] Historical data query templates refer to the validated SQL query scripts and their corresponding business description documents written by technical personnel during the bank's historical operations.

[0112] Business indicator definition files are metadata documents that describe various business indicators of a bank.

[0113] For example, the definition file for the "personal loan non-performing rate" indicator includes information such as indicator name, indicator code, business meaning, calculation formula, statistical period, dimensional attributes, data source, data retrieval logic, and threshold rules. These definition files are stored in formats such as JSON, XML, or Excel, and are used to inject standardized definitions of business indicators into the knowledge base during the cold start phase.

[0114] When the knowledge base self-evolution module is in the continuous learning phase, it extracts new business rules from the dialogue between the user and the AI ​​data question-and-answer module and updates the knowledge base.

[0115] When the knowledge base self-evolution module is in the error correction phase, it receives user feedback on the visualization report and updates the knowledge base according to the correction instructions.

[0116] In this embodiment, by setting up a knowledge base self-evolution module and dividing it into three stages—cold start, continuous learning, and error correction—the system can quickly build a basic knowledge base and continuously optimize it during use. This solves the shortcomings of traditional systems that require manual rule updates and are difficult to adapt to business changes, and realizes dynamic updates and self-improvement of the knowledge base, significantly improving the accuracy and business adaptability of AI question answering.

[0117] The basic concepts have been described above. Obviously, for those skilled in the art, the detailed disclosure above is merely illustrative and does not constitute a limitation of this specification. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this specification. Such modifications, improvements, and corrections are suggested in this specification and therefore remain within the spirit and scope of the exemplary embodiments described herein.

[0118] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this specification are not intended to limit the order of the processes and methods described herein. Although various examples have been discussed in the foregoing disclosure of some embodiments of the invention that are currently considered useful, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the spirit and scope of the embodiments described herein. For example, while the system components described above can be implemented using hardware devices, they can also be implemented solely using software solutions, such as installing the described system on existing servers or mobile devices.

[0119] Similarly, it should be noted that, in order to simplify the description disclosed herein and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of embodiments in this specification may sometimes combine multiple features into a single embodiment, drawing, or description thereof. However, this method of disclosure does not imply that the subject matter of this specification requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of a single embodiment disclosed above.

[0120] In some embodiments, numbers describing the quantity of components and attributes are used. It should be understood that such numbers used in the description of embodiments are modified in some examples with the terms "approximately," "approximately," or "generally." Unless otherwise stated, "approximately," "approximately," or "generally" indicates that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may be changed depending on the characteristics required by individual embodiments. In some embodiments, numerical parameters should take into account specified significant digits and employ a general method of digit reservation. Although the numerical ranges and parameters used to confirm their breadth of range in some embodiments of this specification are approximate values, in specific embodiments, such values ​​are set as precisely as feasible.

[0121] For each patent, patent application, patent application publication, and other material such as articles, books, specifications, publications, and documents referenced in this specification, the entire contents of which are incorporated herein by reference. This excludes historical application documents that are inconsistent with or conflict with the content of this specification, as well as documents that limit the broadest scope of the claims in this specification (currently or subsequently appended to this specification). It should be noted that in the event of any inconsistency or conflict between the descriptions, definitions, and / or terminology used in the supplementary materials to this specification and the content of this specification, the descriptions, definitions, and / or terminology used in this specification shall prevail.

[0122] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A banking intelligent decision-making system driven by both BI and AI, characterized in that: include: The data production layer is used to collect multi-source banking business data, classify and label the multi-source banking business data through a business intelligence agent to generate business indicators, and generate data dashboards corresponding to the business indicators. The data transmission layer is used to distribute the data dashboard to user terminals; The data consumption layer includes a multi-terminal adaptation module and an AI data question-and-answer module. The multi-terminal adaptation module is used to display the data dashboard on different types of user terminals. The AI ​​data question-and-answer module is used to receive natural language queries initiated by users for the business indicators, parse the query intent through multiple artificial intelligence agents, obtain the corresponding business indicators from the data production layer, and generate a visual report. The business action layer is used to associate early warning information with pending tasks, create and assign pending tasks according to preset business rules, monitor the execution status of pending tasks, and obtain a handling conclusion when the execution status is completed. The handling conclusion is then fed back to the data production layer and stored in conjunction with the business indicators that triggered the generation of the early warning information.

2. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 1, characterized in that, The data production layer includes: Data storage unit, used to store multi-source banking business data; The data acquisition unit adopts a distributed architecture to collect multi-source banking business data from multiple source systems of the bank, and stores the collected multi-source banking business data in the data storage unit; The data processing unit classifies and tags the multi-source banking business data through a business intelligence agent to generate business indicators; The dashboard generation unit is used to visualize the business indicators into a data dashboard according to personalized needs. The dashboard generation unit includes a visualization component library and a drag-and-drop interactive interface. The drag-and-drop interactive interface is used to receive drag-and-drop operation instructions from users on the components in the visualization component library, and dynamically adjust the layout and display content of the data dashboard based on the drag-and-drop operation instructions.

3. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 2, characterized in that, The data transmission layer includes a real-time push mode and an on-demand pull mode; The real-time push mode is used to push early warning information or visual reports to the user terminal when abnormal business indicators are detected. The on-demand retrieval mode is used to respond to proactive data query requests initiated by users through the user terminal. Based on the proactive data query request, the corresponding business indicators or data dashboards are obtained from the data production layer, and the results are returned to the user terminal that initiated the request for display.

4. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 3, characterized in that, The AI ​​data question-answering module includes multiple AI agents, which adopt a cascaded processing architecture, including: An intent recognition intelligent agent is used to parse natural language queries initiated by users for the business metrics and identify their query intent, time range, and target metrics. A knowledge retrieval agent is used to retrieve indicator definitions and calculation rules related to the query intent from the enterprise knowledge base; SQL generation agent, used to automatically generate data query statements based on the outputs of the intent recognition agent and the knowledge retrieval agent; A data query agent is used to execute the data query statement and obtain the corresponding business indicators from the data production layer. The results interpretation agent is used to perform trend analysis and business interpretation on the acquired business indicators. The report generation agent integrates the output of the result interpretation agent to generate a visual report containing key conclusions, detailed data, and action recommendations.

5. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 4, characterized in that, The business action layer includes: The early warning and monitoring module is used to continuously monitor business indicators and generate early warning information when the business indicator value triggers a preset threshold. The task linkage module is used to receive the warning information, create task assignments according to the warning type and preset business rules, and assign the task assignments to the corresponding personnel. The task monitoring module is used to monitor the execution status of the pending tasks and to provide upgrade reminders for pending tasks whose execution status has expired and has not been updated. The result feedback module is used to obtain the handling conclusion when the execution status of the pending task is completed, and to feed the handling conclusion back to the data production layer, and bind and store it with the business indicator that triggered the generation of the early warning information.

6. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 5, characterized in that, It also includes a knowledge base self-evolution module, which includes a cold start phase, a continuous learning phase, and an error correction phase. When the knowledge base self-evolution module is in the cold start phase, historical data query templates and business indicator definition files are imported to build the basic knowledge base. When the knowledge base self-evolution module is in the continuous learning phase, it extracts new business rules from the dialogue between the user and the AI ​​data question-and-answer module and updates the knowledge base. When the knowledge base self-evolution module is in the error correction phase, it receives user feedback on the visualization report and updates the knowledge base according to the correction instructions.

7. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 1, characterized in that, The multi-terminal adaptation module includes: Terminal type identification unit, used to obtain the type of terminal device currently logged in by the user; The content adaptive rendering unit is communicatively connected to the terminal type identification unit and is configured to render a complete data dashboard that supports multi-dimensional drill-down when the terminal device type is identified as a web page; and extract business indicators from the data dashboard and perform lightweight rendering according to mobile interaction specifications when the terminal device type is identified as a mobile device. The context synchronization unit is used to synchronize the current query context, filter conditions, and alert information between different terminals when a user switches terminal devices.

8. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 7, characterized in that, The content adaptive rendering unit includes: The rendering strategy experiment pool is used to store multiple candidate rendering strategies preset for different terminal types and user query scenarios; A rendering decision scorer is used to obtain user operation data of different candidate rendering strategies on corresponding terminal types in historical data, and to calculate the experience score of each candidate rendering strategy based on the user operation data. The rendering strategy selection subunit is connected to the terminal type identification unit and the rendering decision scorer. It is used to select the candidate rendering strategy with the highest experience score from the rendering strategy experiment pool as the target rendering strategy according to the current terminal device type. A dynamic rendering executor is used to render the data dashboard according to the target rendering strategy.

9. The banking intelligent decision-making system based on BI and AI dual-drive as described in claim 8, characterized in that, The rendering decision scorer calculates the experience score using the following logic: ； In the formula, Indicates the first Experience scores for candidate rendering strategies on terminal type j; This represents the click-through rate, which is the ratio of valid interactions to total interactions. This represents the page dwell time index, which is a normalized value of the actual dwell time and the baseline time. This represents the task completion rate, which is the ratio of the number of times the target operation was completed to the total number of visits. This represents the early exit rate, which is the ratio of the number of visits with a dwell time below a preset threshold to the total number of visits. , These are the preset weighting coefficients.

Images