An attribution method for business metric anomalies, a program product, an electronic device, and a storage medium

Abstract

This application provides a method, program product, electronic device, and storage medium for attributing anomalies in business metrics, applied in the field of business processing technology. The method includes: receiving a business question regarding anomalies in business metrics; analyzing the business question to determine core parameters for attribution analysis; retrieving business data and / or non-business data associated with the business metrics based on the core parameters to form an analysis context; and generating a target attribution conclusion for the business metric anomaly by reasoning through multiple collaborative agents and a large model based on the analysis context. The method provided in this application can improve the efficiency of attributing anomalies in business metrics, reduce the false positive rate of attributing anomalies in business metrics, and improve the coverage of attribution of anomalies in business metrics.

Description

Technical Field

[0001] This application relates to the field of business processing technology, and more specifically, to an attribution method, program product, electronic device, and storage medium for business indicator anomalies. Background Technology

[0002] Monitoring anomalies in financial business metrics is a crucial step for banks, internet finance institutions, and other organizations to achieve refined operations, real-time risk management, and business decision support. By promptly monitoring and accurately attributing abnormal fluctuations in core business indicators such as loan application numbers, conversion rates, and loan disbursement rates, operations and R&D teams can quickly pinpoint the root causes of problems, formulate effective response strategies, and ensure the steady development of the business. Therefore, establishing a comprehensive anomaly detection system allows for efficient and comprehensive understanding of business data dynamics, further ensuring the stable development of the business.

[0003] Currently, the industry's analysis and attribution of abnormal business metrics mainly relies on a traditional model combining threshold alerts and manual investigation. Specifically, business personnel first set fixed threshold rules based on experience; when the system detects a metric exceeding the threshold, it issues an alert through an office collaboration platform (such as Lark); subsequently, data analysts or operations experts need to intervene, manually performing a series of operations to conduct comprehensive analysis and judgment, and form an attribution conclusion. However, using this traditional model to attribute abnormal business metrics suffers from low efficiency, high false alarm rates, and incomplete attribution coverage. Summary of the Invention

[0004] The purpose of this application is to provide a method, program product, electronic device and storage medium for attributing changes in business indicators, so as to solve the problems of low efficiency, high false alarm rate and incomplete attribution coverage when using the above-mentioned traditional mode to attribute changes in business indicators.

[0005] In a first aspect, embodiments of this application provide an attribution method for business indicator anomalies, comprising: receiving a business question regarding the business indicator anomaly; analyzing the business question to determine core parameters for attribution analysis; retrieving business data and / or non-business data associated with the business indicator based on the core parameters to form an analysis context; and generating a target attribution conclusion for the business indicator anomaly by reasoning through multiple collaborative agents and a large model based on the analysis context.

[0006] The above solution provides a fully automated attribution method. By receiving business issues, intelligently parsing parameters, retrieving data from multiple sources, and utilizing intelligent agents and large models for collaborative reasoning, it achieves end-to-end automation from problem discovery to attribution conclusion generation. Compared to the traditional model in existing technologies that mainly relies on a combination of threshold alarms and manual investigation, the method provided in this application can improve the efficiency of attributing anomalies in business indicators, reduce the false alarm rate of attributing anomalies in business indicators, and improve the coverage of attributing anomalies in business indicators.

[0007] In an optional implementation, analyzing the business problem to determine core parameters for attribution analysis includes: identifying the intent of the business problem; when the intent is for attribution analysis, extracting the core parameters from the business problem based on a business knowledge base, wherein the core parameters include at least one of metrics, dimensions, and time parameters. In the above solution, by identifying user intent and accurately extracting core parameters based on a business knowledge base, accurate understanding and structured transformation of fuzzy natural language queries are achieved, thus laying a reliable foundation for subsequent automated processes, avoiding errors in analysis direction due to misunderstanding of intent, and thereby reducing the false positive rate of attributing changes in business metrics.

[0008] In an optional implementation, after extracting the core parameters from the business problem based on the business knowledge base, the method further includes: completing the missing core parameters based on the historical analysis records of the user who initiated the business problem; and / or, completing the missing core parameters based on a predefined business rule base. In the above scheme, by combining user historical habits with preset business rules to complete the missing parameters, automated parameter completion without manual intervention is achieved, thereby reducing the false alarm rate of attributing anomalies in business metrics. Furthermore, completing the missing parameters can also reduce the false alarm rate of attributing anomalies in business metrics.

[0009] In an optional implementation, after extracting the core parameters from the business problem based on the business knowledge base, the method further includes: performing business rule verification and / or data existence verification on the core parameters. This solution adds an automated verification step for the core parameters. Business rule verification ensures the logical correctness of the matching between indicators and dimensions, while data existence verification avoids invalid or out-of-range queries, thereby reducing the false alarm rate when attributing changes in business indicators.

[0010] In an optional implementation, retrieving business data and / or non-business data associated with the business metrics based on the core parameters includes: determining a first retrieval time window for the business data and a second retrieval time window for the non-business data based on time information in the core parameters, wherein the second retrieval time window is longer than the first retrieval time window; retrieving the business data from a structured database within the first retrieval time window, and / or retrieving the non-business data from an unstructured data source within the second retrieval time window. In the above scheme, since the impact of internal business events is immediate, while the impact of external non-business events is delayed, setting differentiated retrieval time windows for business data and non-business data can improve the retrieval coverage of both explicit and implicit influencing factors, thereby improving the coverage of attribution for changes in business metrics.

[0011] In an optional implementation, the plurality of agents includes a vertical domain agent and / or at least one topic-specific agent, wherein the vertical domain agent provides domain expertise, and the topic-specific agent breaks down the attribution analysis task. In the above scheme, by introducing a vertical domain agent, the professionalism of attributing changes in business metrics is ensured; by introducing a topic-specific agent, complex attribution tasks are logically decomposed, thereby enabling attribution reasoning in complex scenarios and reducing the false alarm rate when attributing changes in business metrics.

[0012] In an optional implementation, the step of generating a target attribution conclusion for the anomaly of the business indicator based on the analysis context, through reasoning by multiple collaborative agents and a large model, includes: decomposing the attribution analysis task into multiple sub-tasks based on the topic agent; for each sub-task, calling the large model to perform reasoning based on the analysis context and the domain expertise provided by the vertical agent to obtain a local attribution conclusion for that sub-task; determining the core attribution event based on the local attribution conclusions of each sub-task, and integrating them to generate the target attribution conclusion. In the above scheme, by using topic agents to decompose the task and collaborating with vertical agents and a large model to perform sub-task reasoning, a step-by-step analysis of attribution reasoning in complex scenarios is achieved. That is, by transforming a fuzzy attribution problem into a series of sub-task reasoning processes, and finally integrating them into a conclusion, the false positive rate of attributing anomalies of business indicators is reduced.

[0013] In an optional implementation, determining the core attribution event based on the local attribution conclusions of each subtask includes: calculating the attribution value score and the relevance score for each related event in the analysis context; and determining the core attribution event based on a weighted fusion score of the attribution value score and the relevance score. In the above scheme, by calculating the attribution value score and the relevance score for each related event and filtering the core attribution event based on the weighted fusion score, the accuracy of filtering the core attribution time is improved, thereby reducing the false positive rate of attributing anomalies in business metrics.

[0014] In an optional implementation, the attribution value score is calculated based on at least one of the following dimensions: the correlation between the occurrence time of the related event and the occurrence time of the business indicator anomaly, the correlation between the related event and the business indicator in terms of business logic, and the consistency of the impact of similar events on historical indicators; and / or, the correlation score is calculated based on at least one of the following dimensions: the matching degree between the related event and the business indicator in terms of business domain, and the correlation degree between the data source of the related event and the data source of the business indicator. In the above scheme, the attribution value score quantifies the impact intensity of the event from multiple dimensions such as time correlation, business logic, and historical impact; the correlation score assesses the basic relevance of the event from the business domain and data source, thereby reducing the false positive rate of attributing business indicator anomalies.

[0015] Secondly, embodiments of this application provide an attribution device for business indicator anomalies, comprising: a receiving module for receiving business questions regarding business indicator anomalies; an analysis module for analyzing the business questions to determine core parameters for attribution analysis; a retrieval module for retrieving business data and / or non-business data associated with the business indicators based on the core parameters to form an analysis context; and an inference module for generating a target attribution conclusion for the business indicator anomalies by inferring from multiple collaborative agents and a large model based on the analysis context.

[0016] The above solution provides a fully automated attribution device that, by receiving business issues, intelligently parsing parameters, retrieving data from multiple sources, and utilizing intelligent agents and large models for collaborative reasoning, achieves end-to-end automation from problem discovery to attribution conclusion generation. Compared to the traditional mode in existing technologies that mainly relies on a combination of threshold alarms and manual investigation, the device provided in this application can improve the efficiency of attributing anomalies in business indicators, reduce the false alarm rate of attributing anomalies in business indicators, and increase the coverage of attributing anomalies in business indicators.

[0017] In an optional implementation, the analysis module is specifically used to: identify the intent of the business question; when the intent is the attribution analysis, extract the core parameters from the business question based on a business knowledge base, wherein the core parameters include at least one of indicators, dimensions, and time parameters. In the above solution, by identifying user intent and accurately extracting core parameters based on a business knowledge base, accurate understanding and structured transformation of fuzzy natural language queries are achieved, thus laying a reliable foundation for subsequent automated processes, avoiding errors in analysis direction due to misunderstanding of intent, and thereby reducing the false positive rate of attributing changes in business indicators.

[0018] In an optional implementation, the attribution device for business indicator anomalies further includes: a completion module, used to complete missing core parameters based on the historical analysis records of the user who initiated the business issue; and / or, to complete missing core parameters based on a predefined business rule base. In the above scheme, by combining user historical habits with preset business rules to complete missing parameters, automated parameter completion without manual intervention is achieved, thereby reducing the false alarm rate for attributing business indicator anomalies. Furthermore, completing missing parameters can also reduce the false alarm rate for attributing business indicator anomalies.

[0019] In an optional implementation, the attribution device for business indicator anomalies further includes a verification module, used to perform business rule verification and / or data existence verification on the core parameters. In the above scheme, an automated verification step for the core parameters is added. Business rule verification ensures the logical correctness of the matching between indicators and dimensions, while data existence verification avoids invalid or out-of-range queries, thereby reducing the false alarm rate when attributing business indicator anomalies.

[0020] In an optional implementation, the retrieval module is specifically used to: determine a first retrieval time window for the business data and a second retrieval time window for the non-business data based on the time information in the core parameters, wherein the second retrieval time window is longer than the first retrieval time window; within the first retrieval time window, retrieve the business data from a structured database, and / or, within the second retrieval time window, retrieve the non-business data from an unstructured data source. In the above scheme, since the impact of internal business events is immediate, while the impact of external non-business events is delayed, by setting differentiated retrieval time windows for business data and non-business data, the retrieval coverage of both explicit and implicit influencing factors can be improved, thereby increasing the coverage of attribution of business indicator anomalies.

[0021] In an optional implementation, the plurality of agents includes a vertical domain agent and / or at least one topic-specific agent, wherein the vertical domain agent provides domain expertise, and the topic-specific agent breaks down the attribution analysis task. In the above scheme, by introducing a vertical domain agent, the professionalism of attributing changes in business metrics is ensured; by introducing a topic-specific agent, complex attribution tasks are logically decomposed, thereby enabling attribution reasoning in complex scenarios and reducing the false alarm rate when attributing changes in business metrics.

[0022] In an optional implementation, the inference module is specifically used for: decomposing the attribution analysis task into multiple sub-tasks based on the topic-specific intelligent agent; for each sub-task, invoking the large model to perform inference based on the analysis context and the domain expertise provided by the vertical domain intelligent agent to obtain a local attribution conclusion for that sub-task; determining the core attribution event based on the local attribution conclusions of each sub-task, and integrating them to generate the target attribution conclusion. In the above scheme, by using the topic-specific intelligent agent to decompose the task and collaborating with the vertical domain intelligent agent and the large model to perform sub-task inference, a step-by-step analysis of attribution inference in complex scenarios is achieved. That is, by transforming a fuzzy attribution problem into a series of sub-task inference processes and finally integrating them into a conclusion, the false alarm rate of attributing changes in business indicators is reduced.

[0023] In an optional implementation, the inference module is further configured to: calculate an attribution value score and a correlation score for each associated event in the analysis context; and determine the core attribution event based on a weighted fusion score of the attribution value score and the correlation score. In the above scheme, by calculating an attribution value score and a correlation score for each associated event and filtering core attribution events based on a weighted fusion score, the accuracy of filtering core attribution events is improved, thereby reducing the false positive rate of attributing changes in business metrics.

[0024] In an optional implementation, the attribution value score is calculated based on at least one of the following dimensions: the correlation between the occurrence time of the related event and the occurrence time of the business indicator anomaly, the correlation between the related event and the business indicator in terms of business logic, and the consistency of the impact of similar events on historical indicators; and / or, the correlation score is calculated based on at least one of the following dimensions: the matching degree between the related event and the business indicator in terms of business domain, and the correlation degree between the data source of the related event and the data source of the business indicator. In the above scheme, the attribution value score quantifies the impact intensity of the event from multiple dimensions such as time correlation, business logic, and historical impact; the correlation score assesses the basic relevance of the event from the business domain and data source, thereby reducing the false positive rate of attributing business indicator anomalies.

[0025] Thirdly, embodiments of this application provide a computer program product, including computer program instructions, which, when read and executed by a processor, perform the attribution method for abnormal business indicators as described in the first aspect.

[0026] Fourthly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus; the processor and the memory communicate with each other via the bus; the memory stores computer program instructions executable by the processor, and the processor can execute the attribution method for abnormal business indicators as described in the first aspect by calling the computer program instructions.

[0027] Fifthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions. When the computer program instructions are executed by a computer, the computer performs the attribution method for abnormal business indicators as described in the first aspect.

[0028] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, embodiments of this application are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A flowchart illustrating an attribution method for business indicator anomalies provided in this application embodiment; Figure 2 An architecture diagram of an attribution method for business indicator anomalies provided in an embodiment of this application; Figure 3 A workflow diagram illustrating an attribution method for business indicator anomalies provided in this application embodiment; Figure 4 This application provides a structural block diagram of an attribution device for abnormal business indicators. Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0031] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0032] Please refer to Figure 1 , Figure 1 A flowchart illustrating an attribution method for abnormal business metrics provided in this application embodiment. This method can, but is not limited to, be executed by an electronic device. Figure 5 The possible structure of this electronic device is shown below; for details, please refer to the following section. Figure 5 The explanation is as follows. Specifically, the attribution methods for the aforementioned business indicator anomalies may include: S101: Receive business-related questions regarding changes in business metrics.

[0033] Business metrics refer to the quantitative measurement of business processes in the financial sector, such as the number of loan and credit applications and the conversion rate of credit disbursements.

[0034] Business metric fluctuations refer to changes in business metrics, typically at the dimensional level. Generally, fluctuations may involve comparing the current period with the baseline period for the same entity, such as a decrease of more than 20% in the current 5-minute traffic of loan applications compared to the average traffic over the same period in the past 1, 7, and 15 days; or, they may involve comparisons between different entities during the same period, such as a surge in credit applications through the wallet entry point compared to the same period last year.

[0035] Business questions refer to natural language queries submitted by users regarding unexpected changes in business metrics, such as "reasons for the decline in credit application conversion rate".

[0036] For example, electronic devices can receive natural language queries input by users through integrated conversational artificial intelligence (AI) interfaces (such as Lark AI). Conversational AI can perform data analysis through natural language dialogue, helping businesses to promptly identify anomalies, pinpoint causes, improve the efficiency of data analysis, and support accurate business decisions. A natural language query refers to a user's inquiry or question using everyday, unstructured natural language.

[0037] S102: Analyze business issues to determine the core parameters used for attribution analysis.

[0038] Attribution refers to looking beyond the surface to understand the underlying causes of changes in business metrics. For example, it might be due to a marketing campaign launched at the wallet entry point to attract users to apply for credit. Analysis, on the other hand, involves interpreting the results of attributing changes in business metrics and providing guidance and suggestions based on the data. For example, it might address the differences in conversion rates resulting from different activities, copywriting, and strategies, and how to better acquire customers.

[0039] Core parameters refer to the key boundary conditions and data retrieval criteria determined through in-depth analysis and structured processing of users' natural language business questions, used to guide subsequent fully automated attribution analysis. For example, core parameters may include metrics (e.g., credit allocation), dimensions (e.g., conversion rate), and time (e.g., this week).

[0040] In step S102 above, the electronic device can perform semantic understanding and structured parsing of the business problem, extracting or deriving the boundary conditions for analysis. One possible approach is that step S102 may specifically include: S201: Identify the intent behind the business problem.

[0041] Intent refers to the fundamental purpose of a user's query; as one implementation method, intent can be classified into attribution analysis, predictive analysis, data query, and report generation.

[0042] For example, electronic devices can input business questions into a large model. The large model, combined with the "financial business analysis scenario classification rules" retrieved by Retrieval-Augmented Generation (RAG), can determine the user's intent and thus clarify the focus of subsequent parameter extraction.

[0043] S202: When the intent is attribution analysis, core parameters are extracted from the business problem based on the business knowledge base. The core parameters include at least one of the following: metrics, dimensions, and time parameters.

[0044] A business knowledge base refers to a collection that stores knowledge such as financial business indicator definitions, dimensional systems, and analysis rules, including structured databases and unstructured data sources. For example, the business knowledge base can be accessed through RAG technology.

[0045] For example, when the intent is attribution analysis, the electronic device can perform parameter extraction based on the identified attribution intent. One implementation method is to extract indicator parameters; another is to extract time parameters; and yet another is to extract dimensional parameters. The electronic device can first use a large model combined with a financial business knowledge base to determine the type of analysis to which the user's need belongs, thus defining the scope for parameter extraction. For instance, if the business question contains keywords such as "cause" or "why," it is identified as attribution analysis; if the business question contains keywords such as "prediction" or "future," it is identified as predictive analysis.

[0046] For example, business common sense knowledge and RAG knowledge base are introduced to assist in the breakdown. For instance, when breaking down the requirement of "declining credit application conversion rate", the "financial business indicator system" can be searched through RAG, and the business rule of "conversion rate = number of applications / number of contacts" can be automatically associated. The indicators (such as credit application conversion rate, number of applications, number of contacts, etc.), dimensions (such as omnichannel / wallet entry), time (such as the default last 7 days) and other parameters can be extracted simultaneously without manual sorting.

[0047] S103: Based on the core parameters, retrieve business data and / or non-business data associated with business metrics to form an analytical context.

[0048] Analytical context refers to a structured information package containing time-series business data, related events, and business knowledge, which is integrated for attribution reasoning.

[0049] Business data refers to internal factors, such as business changes caused by changes in marketing activities, the launch of new products, and a series of operations such as page redesign. Specific indicator information can be located through business document information.

[0050] For example, business data originates from internal business operations and changes, primarily acquired through structured data collection and system integration. Specific paths include: 1. Directly connecting to internal business system interfaces: Real-time or near-real-time collection of operation records such as "marketing campaign launch / decommissioning," "new product feature release," and "page module adjustment" from systems like the operations management platform, product release system, and page redesign management tools. These records typically contain structured information such as clear event names, execution times, and associated business modules, allowing direct identification of corresponding business metrics (e.g., "loan approval application" metric). 2. Parsing business documents and logs: Extracting key event information (e.g., "marketing campaign A launched on the wallet entry point on May 10, 2024") from internal business documents (e.g., operation activity approval forms, product iteration requirement forms) and system operation logs (e.g., backend configuration modification logs) to complete the structured transformation of business events.

[0051] Non-business events refer to external macroeconomic events that impact the credit industry, such as the release of new regulations or changes in the economic market. These events lack specific indicators, and attribution requires converting natural language text into indicator data, thus necessitating a certain level of natural language processing capabilities.

[0052] For example, non-business events focus on external macroeconomic influencing factors, and the acquisition method is centered on unstructured data collection and natural language processing. Specific paths include: 1. Multi-source external information crawling and access: Collecting unstructured text information such as "new credit regulations" from policy release platforms, "economic market fluctuations (such as interest rate adjustments)" from financial information platforms, and "changes in consumption trends" from industry reports through crawling tools or third-party data interfaces. 2. Natural language processing transformation: Since non-business events lack directly related indicator data, the collected text needs to be processed using natural language processing technology, including text cleaning, keyword extraction (e.g., extracting key information such as "credit interest rate adjustment" and "June 1, 2024" from "personal credit interest rate cap adjustment implemented from June 1, 2024"), and event type labeling, transforming the unstructured text into structured event data that can be used for attribution analysis (e.g., "Event Name: Credit Interest Rate Cap Adjustment; Event Time: 2024-06-01; Affected Area: Credit Business").

[0053] One possible approach is to leverage self-explanatory features and combine them with long-term data feedback to continuously refine the Prompt design and improve the accuracy of event attribution. The core role of Prompt rules lies in two main scenarios: 1. Assisting in intent recognition, such as transforming vague requests like "reasons for the decline in credit applications" into structured instructions containing indicators, timeframes, and dimensions, helping to clarify the boundaries of analysis; 2. Standardizing large-scale model attribution reasoning, such as requiring models to output the reasoning process in the order of temporal relevance, business logic matching, and consistency with historical impact, avoiding fragmented conclusions.

[0054] Specifically, firstly, when the large model outputs attribution conclusions, it simultaneously generates Prompt rule application instructions (i.e., self-explanatory), such as noting "This inference did not cover policy-related non-business events because the Prompt did not explicitly require retrieving policy knowledge from documents." Secondly, these instructions are integrated with business feedback (such as the omission of implicit factors discovered through manual verification), and the Prompt rules are adjusted in reverse, for example, by adding constraints such as "It is necessary to call RAG to retrieve policy-related non-business events from Lark documents in the past 30 days," or supplementing examples such as "Policy events need to be associated with the historical impact of credit business indicators." During this process, the RAG knowledge base (providing business knowledge retrieval) and the log system (recording feedback data) in the infrastructure layer provide support for optimization, ensuring that rule iterations align with actual business needs.

[0055] Another possible approach is to use a large model to score the cleaned events in multiple rounds to assess their impact on business metrics. Specifically, attribution value batch scoring quantifies the impact of events, and this score can be further used to optimize Prompt rules. For example, if multiple scoring sessions reveal that "marketing campaign events" have a high attribution accuracy (because the Prompt provides detailed descriptions of the analysis dimensions for this type of event), while "policy events" have a low accuracy (because the Prompt does not clearly define the relationship between policies and metrics), then targeted supplementary rules could be added, such as "policy events need to analyze the time difference between release time and metric changes" and "the overlap between policy coverage areas and business areas," making the Prompt's guidance for different types of events more targeted.

[0056] In step S103 above, the electronic device can automatically and concurrently obtain the required information from multiple data sources based on core parameters. One possible approach is that step S103 may specifically include: S301: Based on the time information in the core parameters, determine the first retrieval time window for business data and the second retrieval time window for non-business data, wherein the second retrieval time window is longer than the first retrieval time window.

[0057] The first retrieval time window refers to a shorter retrieval period set for business data and business events; the second retrieval time window refers to a longer retrieval period set for non-business data and non-business events to cover their delayed impact. For example, electronic devices can use the time of indicator anomaly as a benchmark, setting the first retrieval window for business events and corresponding business data to one day before and one day after the indicator anomaly, and setting the retrieval window for non-business events and corresponding non-business data to seven days before and three days after the indicator anomaly. For example, if the number of credit applications decreased this Wednesday, the business event range is from Tuesday to Thursday of this week, and the non-business event range is from last Wednesday to this Friday.

[0058] S302: Within the first retrieval time window, retrieve business data from a structured database, and / or, within the second retrieval time window, retrieve non-business data from an unstructured data source.

[0059] For example, the electronic device generates and executes an SQL query to obtain time-series data of a specified indicator in the first window from databases such as MySQL / ODPS; the electronic device collects text from external sources through web crawlers or APIs, and performs text cleaning, keyword extraction, and event type labeling through NLP technology to transform it into structured event data.

[0060] S104: Based on the analysis context, multiple agents working collaboratively with a large model perform reasoning to generate target attribution conclusions for anomalies in business metrics.

[0061] An intelligent agent is a software entity capable of autonomously perceiving its environment and making decisions. In financial marketing, it can undertake tasks such as customer service and approval. For example, multiple intelligent agents may include a domain-specific intelligent agent and / or at least one topic-specific intelligent agent, wherein the domain-specific intelligent agent is used to provide domain expertise, and the topic-specific intelligent agent is used to break down attribution analysis tasks.

[0062] Specifically, electronic devices can invoke an agent network composed of vertical domain agents and topic-specific agents to work collaboratively with a large model and perform deep reasoning on the analytical context. One possible approach is that S104 above may specifically include: S401: The attribution analysis task is broken down into multiple sub-tasks based on thematic intelligent agents.

[0063] For example, the topic-based intelligent agent adopts a thinking chain model to break down macro-level problems, such as breaking down "declining conversion rate" into sub-problems such as "reach stage", "application stage", and "loan disbursement stage".

[0064] S402: For each subtask, the large model is invoked to perform inference based on the analysis context and the domain expertise provided by the vertical agent to obtain the local attribution conclusion for that subtask.

[0065] For example, for the application process analysis subtask, the topic-specific agent calls the "credit vertical domain agent" to provide the business rules and common problem patterns for that process, and organizes the analysis context as input, submitting it to the large model. Based on this, the large model performs causal reasoning and outputs local attribution conclusions such as "The application process is abnormal, possibly due to page loading delays or tightened risk control strategies."

[0066] Specifically, this can be achieved by integrating large-scale model attribution reasoning with RAG knowledge enhancement. For example, a COT model using probability analysis and step-by-step reasoning can be adopted to output attribution conclusions for each stage (e.g., "reach is normal, but user open rate is low, suggesting the copywriting does not match user needs"). Furthermore, a "financial business knowledge base" (including business behaviors and strategies) and a "professional knowledge base" (including definitions and business implications of indicators and dimensions) can be invoked, combined with a feedback mechanism that feeds back into the attribution data system, to continuously optimize the accuracy of inference.

[0067] S403: Determine the core attribution event based on the local attribution conclusions of each subtask, and integrate them to generate the target attribution conclusion.

[0068] For example, electronic devices summarize all local attribution conclusions, identify recurring or significant events, determine core attribution events through subsequent quantitative scoring, and finally integrate them into a complete report.

[0069] Furthermore, based on the above embodiments, S403 may specifically include: S501: For each associated event in the analysis context, calculate the attribution value score and the association score for each associated event.

[0070] Attribution value score is a core indicator for quantifying the impact of events on changes in business metrics. One possible approach is to calculate the attribution value score based on at least one of the following dimensions: the correlation between the occurrence time of the related event and the occurrence time of the business metric change, the correlation between the related event and the business metric in terms of business logic, and the consistency of the impact of similar events on historical metrics.

[0071] Among them, time relevance refers to the degree of matching between the time of the event and the time of the indicator change. For example, the smaller the time difference between "the launch of the wallet entry marketing campaign (May 10)" and "the decrease in the number of credit applications (May 11)," the higher the attribution value score. Business logic relevance refers to the degree of matching between the event and the indicator in the business field. For example, "tightening of credit policies" is directly related to "the number of credit applications," while "e-commerce coupon activities" are not related to "the number of credit applications." Historical impact consistency refers to the record of the impact of similar events on historical indicators. For example, "insufficient appeal of marketing campaign copywriting" once led to a 15% decrease in the number of applications. If the current event has the same attributes as that historical event, the attribution value score will be increased.

[0072] For example, firstly, scores are assigned based on a single dimension (0-100 points): time relevance (e.g., 90 points for a 1-day time difference, 60 points for 1-3 days, and 30 points for >3 days), business logic relevance (90 points for direct relevance, 50 points for indirect relevance, and 10 points for irrelevance), and consistency of historical impact (80 points for significant historical impact, 50 points for no historical record, and 20 points for no historical impact). Secondly, the scores are adjusted based on the business rules retrieved from the RAG knowledge base. For example, although "policy-related non-business events" have a large time difference, they are still considered relevant due to policy... The impact has a lag, so the time relevance score can be increased from 30 to 60. Finally, the dimension weights are allocated according to the business scenario. In the attribution scenario, business logic relevance (50%) > time relevance (30%) > historical impact consistency (20%). The formula is: Attribution Value Score = Time Relevance Score × 30% + Business Logic Relevance Score × 50% + Historical Impact Consistency Score × 20%. For example, a marketing campaign has a time relevance score of 80, a business logic relevance score of 90, and a historical impact consistency score of 70, resulting in a final score of 83.

[0073] A relevance score is a metric that assesses the degree of relevance between an event and the current analytical metrics or business scenario. One possible approach is to calculate the relevance score based on at least one of the following dimensions: the degree of matching between the related event and the business metric in the business domain, and the degree of relevance between the data source of the related event and the data source of the business metric.

[0074] Among them, "indicator domain relevance" refers to whether the event belongs to the business domain corresponding to the indicator. For example, when analyzing the number of credit applications, changes in credit policies belong to the credit domain, while e-commerce activities belong to the non-credit domain. "Data source relevance" refers to whether the event data comes from a data source related to the indicator. For example, business events come from the operations management platform, non-business events come from the official website of the central bank's policies, and irrelevant events come from office system logs.

[0075] For example, firstly, the RAG knowledge base and MySQL / ODPS database in the infrastructure layer are called to obtain the business domain definition of the current analysis indicator (e.g., the number of credit applications belongs to the core indicator of credit business) and historical event-indicator correlation records (e.g., the correlation between policy events and credit indicators is 80%). Secondly, the events are scored according to the preset business rules (0-100 points). For example, a complete match of the indicator domain correlation scores 90 points, a partial match scores 50 points, and no match scores 10 points. The data source correlation scores 80 points for directly related data sources, 40 points for indirectly related data sources, and 10 points for unrelated data sources. Finally, the correlation score is calculated using equal weights (50% each), and the formula is: correlation score = indicator domain correlation score × 50% + data source correlation score × 50%. For example, a policy event has an indicator domain correlation score of 90 points and a data source correlation score of 80 points, resulting in a final score of 85 points.

[0076] S502: Determine the core attribution event based on a weighted fusion score of attribution value score and correlation score.

[0077] For example, the weight of the attribution value score can be set to 60% to prioritize the retention of events with strong impact, addressing the pain point of incomplete attribution coverage in existing technologies; the weight of the correlation score can be set to 40% to filter out events with strong impact but irrelevant to the indicators, addressing the pain point of high false positive rates in existing technologies. Specifically, firstly, the comprehensive score of an event can be calculated using a linear weighting formula: Comprehensive Score = Attribution Value Score × 60% + Correlation Score × 40%. For example, an event has an attribution value score of 83 and a correlation score of 85, resulting in a comprehensive score of 83.8 (rounded to 84). Secondly, after merging, events are sorted from high to low based on their comprehensive scores. A dynamic threshold (which can be determined based on historical attribution cases) is set to remove events with scores below the threshold, ensuring that the events ultimately retained have both strong impact and high correlation.

[0078] Furthermore, based on the above embodiments, after S202, the attribution method for business indicator anomalies provided in this application embodiment may further include: S203: Based on the historical analysis records of users who initiated business issues, complete the missing core parameters.

[0079] For example, historical user analysis records in the MySQL / ODPS database at the infrastructure layer can be accessed. If the user's default dimension for analyzing "credit conversion rate" was "wallet entry," this dimension can be automatically supplemented. Specifically, if a user selected "7 days" 8 out of the past 10 analyses of "credit application numbers," the default time for supplementation will be "the last 7 days." If a user analyzed "data from the first week of May 2024" last week, and it is still May, the time will be supplemented to "the second week of May 2024." If 70% of the user's historical analysis queries were based on "channel-wallet entry," the supplemented dimension will be "wallet entry." If a user has frequently analyzed "new user groups" recently, the supplemented user group dimension will be "new users."

[0080] One possible approach is to record the "source of completion (user's Xth historical analysis)" in the background after completion. If the user disagrees with the completion result, they can modify it through the "feedback mechanism." The modification record will be updated synchronously to the historical database to optimize the accuracy of subsequent completions.

[0081] S204: Complete the missing core parameters based on the predefined business rule base.

[0082] For example, when searching for "Financial Business Rules" in Lark documents using RAG, if the user does not specify a time period, the system automatically fills in the parameters for the past 7 days. Specifically, it follows the rule that core indicators are filled in by the past 7 days and non-core indicators by the past 30 days. For example, "credit application count" is filled in as "the past 7 days" and "page views" is filled in as "the past 30 days". If it is the end of the month, it is filled in as "the entire month". It also follows the rule of prioritizing the completion of core dimensions and then the completion of all dimensions. For example, "credit application count" is filled in with "channel - all channels" first. If further subdivision is needed, the three core channels of wallet entry, my entry, and banner entry are filled in. For anomaly analysis scenarios, the "base period time" parameter is additionally filled in, such as the time range of "the past 7 days vs. the same period last week".

[0083] One possible approach is to allow users to confirm and adjust the completed parameters after they have been filled in. Specifically, if the completed parameters conform to a typical scenario, the user can proceed directly to the next step without any further action. If the completed parameters are unique, a pop-up window will appear on the page prompting the user to "adjust?" with a "yes / no" option. If the user selects "yes," they will be redirected to the parameter adjustment page.

[0084] In the above solution, by combining user history and preset business rules to complete missing parameters, automated parameter completion without manual intervention is achieved, thereby reducing the false alarm rate when attributing changes in business metrics. Furthermore, completing missing parameters also reduces the false alarm rate when attributing changes in business metrics.

[0085] Furthermore, based on the above embodiments, after S202, the attribution method for business indicator anomalies provided in this application embodiment may further include: S205: Perform business rule validation and / or data existence validation on core parameters.

[0086] For example, the RAG retrieval metrics and dimension matching rules are called. For instance, "number of credit applications" can be associated with the dimensions of "channels and user groups" but not with the dimension of "product category". If "product category" is extracted, it will be automatically removed and a "dimension mismatch" message will be displayed. The MySQL / ODPS database interface of the infrastructure layer is called to verify whether the extracted parameters have corresponding data. For example, when extracting "number of credit applications in January 2023", if there is no such historical data in the database, "time parameter is invalid" will be marked, guiding the user to adjust the time range.

[0087] The above solution adds a step to automatically verify core parameters. Specifically, business rule verification ensures the logical correctness of the matching between indicators and dimensions, and data existence verification avoids invalid or out-of-range queries, thereby reducing the false alarm rate when attributing anomalies in business indicators.

[0088] Furthermore, based on the above embodiments, since manually writing SQL queries and API call instructions is prone to syntax errors or compliance risks, requiring repeated verification and modification, and lacking an automated optimization mechanism, a large model is introduced to optimize the instructions. Specifically, this involves filtering out illegal instructions such as "querying private data," for example, automatically deleting sensitive fields such as "user's mobile phone number" from SQL; and adjusting parameter granularity, for example, optimizing "real-time query" into "5-minute granularity aggregate query."

[0089] Furthermore, based on the above embodiments, since manual data retrieval requires logging into the MySQL database to query business data, the RAG knowledge base to query business rules, and Quick BI to generate charts, followed by manual data integration, and cannot synchronously correlate event data, resulting in incomplete attribution coverage, an architecture of multi-agent and multi-source data collaborative invocation is introduced to achieve fully automated retrieval and task execution. Specifically, it automatically triggers MySQL to query indicator data, RAG to query business knowledge, and +AGENT to decompose sub-tasks without requiring manual system switching; while retrieving business data, it simultaneously retrieves relevant events from the event system, providing scenario context for subsequent attribution.

[0090] Please refer to Figure 2 , Figure 2 The architecture diagram of an attribution method for business indicator anomalies provided in this application embodiment includes an infrastructure layer, a core layer, and a workflow set.

[0091] Specifically, the infrastructure layer is the foundation of the entire architecture, providing basic support for computing power, knowledge retrieval, and data storage. It includes AI and tool components, and data storage. For example, AI and tool components may include Deepseek (a large model responsible for complex attribution reasoning and logical analysis), RAG (Retrieval Enhancement Generation, providing precise retrieval of business knowledge and metric definitions), AGENT (intelligent agents responsible for task decomposition, tool invocation, and process execution), and Quick BI (a visualization tool used for charting the final analysis results). Data storage may include MySQL and ODPS databases (for storing structured information such as business data, logs, and metrics), and Lark Docs (for storing unstructured knowledge, supplementing the RAG knowledge base).

[0092] The core layer is the brain of the architecture, responsible for user intent understanding, task execution, and knowledge fusion. It includes three core modules: workflow invocation, execution planning, and configuration capabilities. Specifically, intent recognition in workflow invocation involves: inputting a "user query," using the Prompt project (combining historical records and samples), and outputting the target workflow. Historical data and sample data in workflow invocation can be used to optimize the accuracy of intent recognition, making the system more aligned with business habits.

[0093] The execution plan specifically includes: First, parsing the core parameters and transforming user requirements into executable core parameters; second, attempting to fill in any missing parameters; then, generating execution commands and outputting executable instructions such as interface calls and SQL queries; and finally, executing tasks and retrieving data, calling the tools and databases in the infrastructure layer to obtain the raw data required for analysis.

[0094] For example, the event system in the core layer can be used to integrate business events such as "operational strategies, marketing events, and user behavior" to provide context for attribution; large model reasoning can be used to output attribution conclusions based on events, data, and RAG knowledge, and to normalize and integrate multi-dimensional conclusions. Configuration capabilities support Prompt configuration, data source configuration, event configuration, and process configuration, making the architecture flexible.

[0095] Workflow collections serve as entry points for business scenarios, covering attribution, prediction, data retrieval, follow-up inquiries, early warning, and business reporting. Among these, the attribution workflow is the core, encompassing problem discovery (anomaly detection), problem localization (core dimension / metric breakdown, thematic positioning, data linkage, and event analysis), and problem resolution (outputting conclusions and implementation strategies).

[0096] Please refer to Figure 3 , Figure 3This application provides a workflow diagram for an attribution method for business indicator anomalies. Specifically, the workflow may include: user input of a query; enhancement of the prompt based on historical user data to identify query intent (i.e., referencing past analysis habits to make intent identification more accurate), determining which workflow category the user's need belongs to: attribution, prediction, query volume, or report; constructing prompt words using common-sense knowledge such as events and business dimensions, enabling the model to identify specific parameters to extract core parameters, and simultaneously inferring the required interfaces and data tables based on the extracted intent, time, indicator, and dimension parameters; vertical domain agents providing professional knowledge support within their respective domains, while multi-topic agents based on "topic data..." Following the logic of "dimensional / indicator decomposition," "topic thinking chain arrangement and analysis process," and "indicator + dimension trend anomaly analysis," the problem is subdivided and decomposed, that is, the analysis conclusions of the topic agent are based on the core dimension / indicator decomposition; through Deepseek attribution inference, probability analysis is generated, a financial business knowledge base (including business behavior, business strategies, etc.) is constructed, and a professional knowledge base (including the definition and business meaning of indicators and dimensions) is constructed, that is, Deepseek attribution analysis is carried out (combining reasoning, data, and knowledge base integration), and the feedback mechanism of the attribution data system is combined to continuously optimize the accuracy of reasoning; conclusions with the same root cause are normalized and a summary of analysis ideas is output; and a visualization presentation is constructed.

[0097] Therefore, firstly, the financial business attribution analysis architecture that integrates multi-agent and retrieval enhancement in the embodiments of this application innovatively designs a hierarchical collaborative architecture of vertical domain agent, topic agent, RAG and large model. Among them, the vertical domain agent focuses on knowledge of financial sub-domains, the topic agent is responsible for the execution of specific analysis scenario tasks, the RAG ensures accurate retrieval of business knowledge, and the large model completes complex logical reasoning, thus realizing an organic combination of domain knowledge professionalism and task execution refinement.

[0098] Secondly, the three-layer configurable technical architecture in this application proposes a three-layer architecture consisting of an infrastructure layer, a core layer, and a workflow set. The infrastructure layer integrates technologies such as large models, RAG, and intelligent agents, as well as multi-source data storage. The core layer realizes intelligent decision-making and task scheduling and supports flexible configuration of prompts, data sources, events, and processes. The workflow set directly undertakes multiple business scenarios, combining technical scalability and business agility.

[0099] Third, the fully automated attribution method for financial business based on the thinking chain in this application eliminates the fully automated process of user query intent recognition, core parameter extraction, multi-agent collaborative task decomposition, large model COT reasoning, RAG knowledge fusion, conclusion normalization and visualization. It realizes the interpretability of the attribution reasoning process through the thinking chain model and solves the problem of untraceable conclusions in complex attribution scenarios of financial business.

[0100] Fourth, the attribution reasoning method driven by multi-source knowledge and data in the embodiments of this application integrates financial business knowledge base (business behavior, strategies, etc.), professional knowledge base (indicator / dimensional definition, etc.) and multi-source business data (structured database, unstructured document) to form an attribution reasoning mechanism driven by knowledge and data, ensuring that the attribution conclusions are both consistent with business logic and supported by data.

[0101] In summary, compared with existing business anomaly handling solutions, the business indicator anomaly attribution method provided in this application embodiment comprehensively solves the three core bottlenecks of inefficient manual analysis, coarse threshold judgment, and incomplete attribution coverage by deeply integrating RAG, AI Agent, mind chain, and large model.

[0102] In addressing the inefficiency of manual analysis, existing solutions rely on experts to connect multiple aspects of product, technology, data, and operations, with an average investigation time exceeding 4 hours. The attribution method for business indicator anomalies provided in this application's embodiments automates the entire process from anomaly discovery to conclusion output, reducing investigation time from hours to minutes, improving efficiency by over 90%. At the same time, by incorporating domain knowledge into vertical domain agents and topic agents, it significantly reduces reliance on experts.

[0103] In addressing the problem of crude threshold determination, existing solutions with fixed thresholds are prone to high false alarms due to periodic fluctuations. The attribution method for business indicator anomalies provided in this application combines business event systems, historical data, and large-scale model reasoning to dynamically adjust anomaly determination thresholds. It can also perform multi-dimensional analysis by associating indicators, dimensions, and events to accurately identify real anomalies.

[0104] In addressing the issue of attribution coverage, existing solutions rely on manual analysis with an identification rate of less than 35% for implicit factors such as policy changes. The attribution method for business indicator anomalies provided in this application retrieves policy documents and business rule bases using RAG, integrates financial business knowledge bases, professional knowledge bases, and multi-source data, and achieves full-dimensional attribution of explicit and implicit factors. This improves the identification rate of implicit factors and provides more efficient, accurate, and comprehensive intelligent support for handling financial business anomalies.

[0105] Please refer to Figure 4 , Figure 4This application provides a structural block diagram of an attribution device for business indicator anomalies. The attribution device 600 includes: a receiving module 601 for receiving business questions regarding business indicator anomalies; an analysis module 602 for analyzing the business questions to determine core parameters for attribution analysis; a retrieval module 603 for retrieving business data and / or non-business data associated with the business indicator based on the core parameters to form an analysis context; and a reasoning module 604 for generating a target attribution conclusion for the business indicator anomalies based on the analysis context, through multiple collaborative agents and a large model.

[0106] The above solution provides a fully automated attribution device that, by receiving business issues, intelligently parsing parameters, retrieving data from multiple sources, and utilizing intelligent agents and large models for collaborative reasoning, achieves end-to-end automation from problem discovery to attribution conclusion generation. Compared to the traditional mode in existing technologies that mainly relies on a combination of threshold alarms and manual investigation, the device provided in this application can improve the efficiency of attributing anomalies in business indicators, reduce the false alarm rate of attributing anomalies in business indicators, and increase the coverage of attributing anomalies in business indicators.

[0107] Furthermore, based on the above embodiments, the analysis module 602 is specifically used to: identify the intent of the business problem; when the intent is the attribution analysis, extract the core parameters from the business problem based on the business knowledge base, wherein the core parameters include at least one of indicators, dimensions, and time parameters.

[0108] In the above solution, by identifying user intent and accurately extracting core parameters based on the business knowledge base, the solution achieves accurate understanding and structured transformation of fuzzy natural language queries, thus laying a reliable foundation for subsequent automated processes, avoiding errors in analysis direction due to misunderstanding of intent, and reducing the false alarm rate of attributing anomalies in business indicators.

[0109] Furthermore, based on the above embodiments, the attribution device 600 for abnormal business indicators further includes: a completion module, used to complete the missing core parameters based on the historical analysis records of the user who initiated the business problem; and / or, to complete the missing core parameters based on a predefined business rule base.

[0110] In the above solution, by combining user history and preset business rules to complete missing parameters, automated parameter completion without manual intervention is achieved, thereby reducing the false alarm rate when attributing changes in business metrics. Furthermore, completing missing parameters also reduces the false alarm rate when attributing changes in business metrics.

[0111] Furthermore, based on the above embodiments, the attribution device 600 for abnormal business indicators also includes: a verification module, used to perform business rule verification and / or data existence verification on the core parameters.

[0112] The above solution adds a step to automatically verify core parameters. Specifically, business rule verification ensures the logical correctness of the matching between indicators and dimensions, and data existence verification avoids invalid or out-of-range queries, thereby reducing the false alarm rate when attributing anomalies in business indicators.

[0113] Furthermore, based on the above embodiments, the retrieval module 603 is specifically used to: determine a first retrieval time window for the business data and a second retrieval time window for the non-business data according to the time information in the core parameters, wherein the second retrieval time window is longer than the first retrieval time window; within the first retrieval time window, retrieve the business data from the structured database, and / or, within the second retrieval time window, retrieve the non-business data from the unstructured data source.

[0114] In the above scheme, since the impact of internal business events is immediate while the impact of external non-business events is delayed, by setting different retrieval time windows for business data and non-business data, the retrieval coverage of both explicit and implicit influencing factors can be improved, thereby increasing the coverage of attribution of business indicator anomalies.

[0115] Furthermore, based on the above embodiments, the plurality of intelligent agents include a vertical domain intelligent agent and / or at least one topic intelligent agent, wherein the vertical domain intelligent agent is used to provide domain expertise, and the topic intelligent agent is used to break down the attribution analysis task.

[0116] In the above solution, the introduction of vertical domain intelligent agents ensures the professionalism of attributing changes in business indicators; the introduction of topic-specific intelligent agents enables the logical decomposition of complex attribution tasks, thereby enabling attribution reasoning in complex scenarios and reducing the false alarm rate of attributing changes in business indicators.

[0117] Furthermore, based on the above embodiments, the inference module 604 is specifically used for: decomposing the attribution analysis task into multiple sub-tasks based on the topic agent; for each sub-task, calling the large model to perform inference based on the analysis context and the domain expertise provided by the vertical agent to obtain the local attribution conclusion of the sub-task; determining the core attribution event based on the local attribution conclusion of each sub-task, and integrating them to generate the target attribution conclusion.

[0118] In the above scheme, the task is decomposed by the topic intelligent agent and the vertical domain intelligent agent and the large model are coordinated to perform sub-task reasoning, thereby realizing the step-by-step analysis of attribution reasoning in complex scenarios. In other words, by transforming a fuzzy attribution problem into a series of sub-task reasoning processes and finally integrating them into a conclusion, the false alarm rate of attributing anomalies in business indicators is reduced.

[0119] Furthermore, based on the above embodiments, the inference module 604 is also used to: calculate the attribution value score and the relevance score of each associated event for each associated event in the analysis context; and determine the core attribution event based on the weighted fusion score of the attribution value score and the relevance score.

[0120] In the above scheme, by calculating the attribution value score and the relevance score for each related event, and filtering the core attribution events based on the weighted fusion score, the accuracy of filtering the core attribution time is improved, thereby reducing the false alarm rate of attributing anomalies in business indicators.

[0121] Furthermore, based on the above embodiments, the attribution value score is calculated based on at least one of the following dimensions: the correlation between the occurrence time of the related event and the occurrence time of the business indicator anomaly, the correlation between the related event and the business indicator in business logic, and the consistency of the impact of similar events on historical indicators; and / or, the correlation score is calculated based on at least one of the following dimensions: the matching degree between the related event and the business indicator in the business domain, and the correlation between the data source of the related event and the data source of the business indicator.

[0122] In the above scheme, the attribution value score quantifies the impact intensity of events from multiple dimensions such as time correlation, business logic, and historical impact; the correlation score assesses the basic relevance of events from the business domain and data source, thereby reducing the false alarm rate of attributing anomalies in business indicators.

[0123] Please refer to Figure 5 , Figure 5 This application provides a structural block diagram of an electronic device 700, which includes at least one processor 701, at least one communication interface 702, at least one memory 703, and at least one communication bus 704. The communication bus 704 enables direct communication between these components, the communication interface 702 facilitates signaling or data communication with other node devices, and the memory 703 stores machine-readable instructions executable by the processor 701. When the electronic device 700 is running, the processor 701 communicates with the memory 703 via the communication bus 704. When the machine-readable instructions are invoked by the processor 701, the aforementioned attribution method for abnormal business indicators is executed.

[0124] For example, the processor 701 in this embodiment of the application can read a computer program from the memory 703 via the communication bus 704 and execute the computer program to implement the following method: receiving a business question about an anomaly in a business indicator; analyzing the business question to determine core parameters for attribution analysis; retrieving business data and / or non-business data associated with the business indicator based on the core parameters to form an analysis context; and generating a target attribution conclusion for the anomaly in the business indicator by reasoning with multiple agents working collaboratively and a large model based on the analysis context.

[0125] The processor 701 comprises one or more, and can be an integrated circuit chip with signal processing capabilities. The processor 701 can be a general-purpose processor, including a Central Processing Unit (CPU), a Microcontroller Unit (MCU), a Network Processor (NP), or other conventional processors; it can also be a special-purpose processor, including a Neural-network Processing Unit (NPU), a Graphics Processing Unit (GPU), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Furthermore, when there are multiple processors 701, some can be general-purpose processors, and others can be special-purpose processors.

[0126] The memory 703 includes one or more, which may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0127] Understandable. Figure 5 The structure shown is for illustrative purposes only; the electronic device 700 may also include components that are more advanced than those shown. Figure 5 The more or fewer components shown, or having the same Figure 5 The different configurations shown. Figure 5 The components shown can be implemented using hardware, software, or a combination thereof. In the embodiments of this application, electronic device 700 can be, but is not limited to, physical devices such as desktop computers, laptops, smartphones, smart wearable devices, and in-vehicle devices, or virtual devices such as virtual machines. Furthermore, electronic device 700 is not necessarily a single device; it can be a combination of multiple devices, such as a server cluster, etc.

[0128] This application also provides a computer program product, including a computer program stored on a computer-readable storage medium. The computer program includes computer program instructions. When the computer program instructions are executed by a computer, the computer is able to perform the steps of the attribution method for business indicator anomalies described in the above embodiments.

[0129] This application also provides a computer-readable storage medium that stores computer program instructions. When the computer program instructions are executed by a computer, the computer performs the attribution method for abnormal business indicators described in the foregoing method embodiments.

[0130] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0131] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0132] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0133] It should be noted that if the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0135] The above description is merely an embodiment of this application and is not intended to limit the scope of protection 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 protection of this application.

Claims

1. An attribution method for abnormal changes in business indicators, characterized in that, include: Receive business-related inquiries regarding changes in business metrics; Analyze the business problem to determine the core parameters used for attribution analysis; Based on the core parameters, retrieve business data and / or non-business data associated with the business metrics to form an analytical context; Based on the aforementioned analytical context, multiple agents working collaboratively with a large model perform reasoning to generate target attribution conclusions for the anomalies in the aforementioned business metrics.

2. The attribution method for business indicator anomalies according to claim 1, characterized in that, The analysis of the business problem aims to determine the core parameters used for attribution analysis, including: Identify the intent behind the business problem; When the intent is the attribution analysis, the core parameters are extracted from the business problem based on the business knowledge base, wherein the core parameters include at least one of indicators, dimensions, and time parameters.

3. The attribution method for business indicator anomalies according to claim 2, characterized in that, After extracting the core parameters from the business problem based on the business knowledge base, the method further includes: Based on the historical analysis records of the users who initiated the aforementioned business issues, the missing core parameters are completed; and / or, The missing core parameters are completed based on a predefined business rule base.

4. The attribution method for business indicator anomalies according to claim 2, characterized in that, After extracting the core parameters from the business problem based on the business knowledge base, the method further includes: Perform business rule verification and / or data existence verification on the core parameters.

5. The attribution method for business indicator anomalies according to claim 1, characterized in that, The step of retrieving business data and / or non-business data associated with the business metrics based on the core parameters includes: Based on the time information in the core parameters, a first retrieval time window for the business data and a second retrieval time window for the non-business data are determined, wherein the second retrieval time window is longer than the first retrieval time window; Within the first retrieval time window, the business data is retrieved from the structured database, and / or, within the second retrieval time window, the non-business data is retrieved from the unstructured data source.

6. The attribution method for business indicator anomalies according to claim 1, characterized in that, The plurality of agents includes a vertical domain agent and / or at least one topic agent, wherein the vertical domain agent is used to provide domain expertise and the topic agent is used to break down the attribution analysis task.

7. The attribution method for business indicator anomalies according to claim 6, characterized in that, Based on the analytical context, the process involves multiple agents working collaboratively with a large model to infer and generate target attribution conclusions for the anomalies in the business metrics, including: Based on the topic-specific intelligent agent, the attribution analysis task is broken down into multiple sub-tasks; For each subtask, the large model is invoked to perform inference based on the analysis context and the domain expertise provided by the vertical domain agent to obtain the local attribution conclusion for that subtask; The core attribution event is determined based on the local attribution conclusions of each subtask, and then integrated to generate the target attribution conclusion.

8. The attribution method for business indicator anomalies according to claim 7, characterized in that, The determination of the core attribution event based on the local attribution conclusions of each sub-task includes: For each associated event in the analysis context, calculate the attribution value score and the relevance score for each associated event; The core attribution event is determined based on a weighted fusion score of the attribution value score and the correlation score.

9. The attribution method for business indicator anomalies according to claim 8, characterized in that, The attribution value score is calculated based on at least one of the following dimensions: the correlation between the occurrence time of the related event and the occurrence time of the business indicator change, the correlation between the related event and the business indicator in business logic, and the consistency of the impact of similar events on historical indicators. And / or, The correlation score is calculated based on at least one of the following dimensions: the degree of matching between the related event and the business indicator in the business domain, and the degree of correlation between the data source of the related event and the data source of the business indicator.

10. A computer program product, characterized in that, It includes computer program instructions, which, when read and executed by a processor, perform the attribution method for business indicator anomalies as described in any one of claims 1-9.

11. An electronic device, characterized in that, include: Processor, memory, and bus; The processor and the memory communicate with each other via the bus; The memory stores computer program instructions that can be executed by the processor, and the processor can invoke the computer program instructions to execute the attribution method for business indicator anomalies as described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a computer, cause the computer to perform the attribution method for business indicator anomalies as described in any one of claims 1-9.

Images