Index transaction attribution analysis method, electronic equipment, storage medium and product

Abstract

The invention provides an index transaction attribution analysis method, electronic equipment, a storage medium and a product, and belongs to the technical field of artificial intelligence. The method comprises the following steps: in response to an attribution analysis request for transaction of a target index, querying target data related to the target index; calling a plurality of agents, and analyzing the target data from a plurality of dimensions to obtain abnormal data under the plurality of dimensions; inputting the abnormal data under the plurality of dimensions into an attribution reasoning model, and outputting a plurality of candidate root causes causing the transaction of the target index; the multiple candidate root causes, the abnormal data under the multiple dimensions and data of a preset knowledge base are input into an attribution analysis model, an attribution analysis result of the target index transaction is output, and the preset knowledge base is used for storing description data of different indexes and dimensions. The method does not depend on a traditional reasoning algorithm, and the attribution analysis result reasoned by means of the model is more comprehensive and more accurate.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an indicator change attribution analysis method, electronic equipment, storage medium and product. Background Art

[0002] Attribution analysis of metric fluctuations involves using quantitative and qualitative methods to determine the causes of unusual changes in metrics. In other words, it involves identifying the key factors influencing these changes. For example, in the e-commerce sector, if a platform's GMV (Gross Merchandise Volume) increases by 5% today compared to yesterday, attribution analysis can identify the cause of this increase and enable appropriate adjustments and optimizations to the platform's services, thereby increasing user conversion rates and revenue.

[0003] The attribution analysis method of indicator changes in related technologies is: obtain the query request input by the user for the change of the target indicator; input the query request into the intention recognition model, and output the user's intention as attribution analysis; based on the attribution analysis intention, use traditional reasoning algorithms (such as contribution algorithms) to infer the root cause of the target indicator abnormality; input the inferred root cause of the target indicator abnormality into the content generation model, and output the attribution analysis result of the target indicator.

[0004] However, due to the limitations of traditional reasoning algorithms themselves, the attribution analysis results of related technologies are not comprehensive or accurate enough. Summary of the Invention

[0005] The present invention provides an indicator change attribution analysis method, electronic device, storage medium, and product. This method does not rely on traditional inference algorithms, and the attribution analysis results derived by model inference are more comprehensive and accurate. The technical solution is as follows:

[0006] In a first aspect, a method for attribution analysis of indicator changes is provided, the method comprising:

[0007] In response to an attribution analysis request for a target indicator change, querying target data related to the target indicator;

[0008] Calling multiple intelligent agents to analyze the target data from multiple dimensions to obtain abnormal data in multiple dimensions;

[0009] Inputting the abnormal data under the multiple dimensions into the attribution reasoning model to output multiple candidate root causes that lead to the abnormal changes in the target indicators;

[0010] The multiple candidate root causes, the abnormal data under the multiple dimensions and the data of the preset knowledge base are input into the attribution analysis model, and the attribution analysis results of the target indicator changes are output. The preset knowledge base is used to store descriptive data and field-related data of different indicators and dimensions.

[0011] In a second aspect, a device for analyzing attribution of indicator changes is provided, the device comprising:

[0012] a query module, configured to query target data related to the target indicator in response to an attribution analysis request for a change in the target indicator;

[0013] An analysis module is used to call multiple agents to analyze the target data from multiple dimensions to obtain abnormal data in multiple dimensions;

[0014] A first input-output module is configured to input the abnormal data in the multiple dimensions into an attribution reasoning model and output multiple candidate root causes that lead to the abnormal change of the target indicator;

[0015] The second input-output module is used to input the multiple candidate root causes, the abnormal data under the multiple dimensions and the data of the preset knowledge base into the attribution analysis model, and output the attribution analysis results of the target indicator changes. The preset knowledge base is used to store descriptive data and field-related data of different indicators and dimensions.

[0016] In a third aspect, an electronic device is provided, comprising a processor and a memory; the memory stores at least one program code; the at least one program code is used to be called and executed by the processor to implement the indicator change attribution analysis method described in the first aspect.

[0017] In a fourth aspect, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, and when the at least one computer program is executed by a processor, the indicator change attribution analysis method described in the first aspect can be implemented.

[0018] In a fifth aspect, a computer program product is provided, which includes a computer program, and when the computer program is executed by a processor, it can implement the indicator change attribution analysis method described in the first aspect.

[0019] The beneficial effects of the technical solution provided by the embodiments of the present application are:

[0020] The embodiment of the present application uses the powerful reasoning ability of artificial intelligence itself to build an intelligent decision-making system driven by artificial intelligence as the core. When performing attribution analysis of indicator changes with the help of this decision-making system, in response to the attribution analysis request for the change of the target indicator, the target data related to the target indicator is queried, and multiple intelligent agents are called to analyze the target data from multiple dimensions to obtain abnormal data under multiple dimensions. The abnormal data under multiple dimensions are then input into the attribution reasoning model, and a variety of candidate root causes that lead to the change of the target indicator are output. Then, the multiple candidate root causes, the abnormal data under multiple dimensions, and the data of the preset knowledge base are input into the attribution analysis model, and the attribution analysis results of the change of the target indicator are output. The method provided in the embodiment of the present application inputs the original objective abnormal data, the knowledge base, and the candidate root causes inferred by the attribution reasoning model into the attribution analysis model for reasoning verification. The analysis process is more scientific, and the attribution analysis results are more comprehensive and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0022] Figure 1 This is an architectural diagram of an indicator change attribution analysis method provided in an embodiment of the present application;

[0023] Figure 2 This is a flow chart of an indicator change attribution analysis method provided by an embodiment of the present application;

[0024] Figure 3 This is a flow chart of an indicator change attribution analysis method provided by an embodiment of the present application;

[0025] Figure 4 This is an indicator change attribution analysis device provided by an embodiment of the present application;

[0026] Figure 5 A structural block diagram of an electronic device provided by an exemplary embodiment of the present application is shown. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0028] It should be understood that the terms "each," "plurality," and "any" used in the embodiments of this application include two or more, "each" refers to each of the corresponding plurality, and "any" refers to any one of the corresponding plurality. For example, if a plurality of words includes 10 words, "each" refers to each of the 10 words, and "any" refers to any one of the 10 words.

[0029] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0030] Before executing the embodiments of the present application, the terms involved in the embodiments of the present application are first explained.

[0031] LLM (Large Language Model) refers to a deep learning model trained using large amounts of text data, enabling it to generate natural language text or understand the meaning of linguistic text. LLM can provide in-depth knowledge and language production on a variety of topics by training on large datasets. Its core concept is to learn the patterns and structure of natural language through large-scale unsupervised training, to a certain extent simulating the human language cognition and generation process.

[0032] Retrieval Augmented Generation (RAG) combines traditional retrieval methods with natural language generation. RAG retrieves relevant documents from a database and uses this information to generate answers or text.

[0033] Long context refers to the ability of large language models to understand and generate long contextual information when processing input text. In natural language processing (NLP), context length refers to the maximum length of a complete sentence that the model can store during inference.

[0034] The context window refers to the amount of text a model can accept when generating or understanding language, or the number of tokens a model can process. The context window determines the amount of contextual information a model can consider during the generation process, helping the model generate coherent and relevant text without the cluttered or irrelevant output that comes from excessive context.

[0035] Data parsing is the process of converting one data format into another readable format. Specifically, it involves analyzing the relationships between the various components of a given data set. After parsing the data, it can be formatted in a preferred format (e.g., JSON). For example, data in HTML (Hypertext Markup Language) format can be converted to a more understandable form using a parser.

[0036] Data ingestion is the process of acquiring data and importing it into a system or storage facility. This acquired data can be used directly, processing and analyzing it immediately to meet immediate needs. Alternatively, it can be stored in a database for later access and processing as needed.

[0037] An agent is an intelligent system or entity that can perceive its environment, make autonomous decisions, and take actions to achieve specific goals. It can access large language models and tools, develop and execute plans, and possess memory to store historical reasoning results.

[0038] Chain-of-Thought (CoT) refers to the process of breaking down complex logical problems into a complete thought process through a series of logically related thoughts. CoT is used to improve the reasoning capabilities of large language models by encouraging them to output intermediate reasoning steps (such as step-by-step thought processes) before generating the final answer, thereby improving the accuracy of complex tasks. For example, when solving mathematical problems or logical reasoning, the model can reduce errors by displaying the chain of reasoning.

[0039] The existing attribution analysis methods for indicator changes only rely on artificial intelligence to perform processing such as intent recognition, parameter extraction, and text polishing, and do not perform reasoning based on the reasoning ability of artificial intelligence itself. The attribution analysis results obtained are solidified, lack reasoning and inspiration, and are not comprehensive or accurate enough.

[0040] The embodiment of this application upgrades the artificial intelligence attribution paradigm, leveraging the powerful reasoning capabilities of artificial intelligence itself to build an intelligent decision-making system driven by artificial intelligence. By inputting the original objective abnormal data, knowledge base, and candidate root causes derived from the attribution reasoning model into the attribution analysis model for reasoning verification, the analysis process is more scientific and the attribution analysis results are more comprehensive and accurate. The innovative features of the embodiment of this application include:

[0041] First, upgrade the AI-driven model: from the existing "AI-assisted" to "AI-led", using large language models for active reasoning to replace the traditional passive response model.

[0042] The second point is the mutual verification of intelligent suggestions and data: using large language models to give possible hypotheses, and then verifying them with the help of real data, it not only provides innovative ideas, but also filters out reliable conclusions through real data, solving the problems of insufficient inspiration and inaccurate attribution analysis results.

[0043] Third, a knowledge base is established based on specific application scenarios. This knowledge base covers various knowledge in the field, allowing the large language model to better understand industry characteristics and thus provide more reliable attribution analysis results.

[0044] Figure 1 A system architecture diagram of an indicator change attribution analysis method provided by an embodiment of the present application is shown. Figure 1 The system has multiple capabilities including front-end display, product capabilities, supporting facilities, workflow collection, platform unified architecture and underlying capabilities.

[0045] Among them, the front-end display can display an interactive dialogue page on the web page, which displays a robot for human-computer interaction, dialogue content, and warning information.

[0046] Product capabilities include front-end services, system monitoring, and engineering capabilities. Front-end services include enriching the card styles displayed on the front-end, providing shortcut commands and question templates, and sharing screenshots. System monitoring includes permission management, data security, log storage, and cached responses. Engineering capabilities include subscription management, anomaly detection, startup alerts, conversation start / stop, and follow-up inquiries.

[0047] Supporting facilities include knowledge bases and databases. Knowledge bases provide relevant knowledge within specific fields, synchronize data, manage knowledge operations, and provide update and maintenance mechanisms. Databases store various tagged data and enable OLAP (Online Analytical Processing) calculations and use algorithms to generate predictions based on stored data. Databases are also scalable, allowing for expanded data metrics and dimensions.

[0048] The workflow collection includes the identified intents and workflows for these intents. Identified intents include attribution analysis, forecasting, table search / number querying, follow-up, early warning, and operational reporting. Taking attribution analysis as an example, the workflow for attribution analysis includes steps such as problem discovery, problem location, problem resolution, and strategic recommendations. Problems can be identified by performing anomaly detection on data. Problems can be located by breaking down core dimensions / indicators, identifying specific topics, and analyzing the current state of data. Problems can be resolved by modeling assumptions, verifying them, and consolidating conclusions. Strategic recommendations can be implemented through strategic implementation.

[0049] The framework layer of the platform's unified architecture can provide streaming engines, configuration prompts, code management, deployment tools, etc. The underlying capabilities include large language models, CoT / Agent, RAG, data engines, etc. With the help of the framework layer and underlying capabilities, the indicator change attribution analysis process includes intent recognition, data retrieval, and root cause reasoning. Among them, the intent recognition process includes: obtaining the query input by the user, and based on the query input by the user and the prompt project (including historical records, few samples, workflow collections, etc.), using the large language model to identify the user's intent. The data retrieval process includes: taking the identified intent as an example for attribution analysis, parsing the core parameters from the query input by the user. If the core parameters parsed from the query input by the user are insufficient, you can try to fill in the parameters, and then generate executable SQL (Structured Query Language) based on the filled core parameters, call the search engine for data retrieval, and obtain the target data. The root cause inference process involves summarizing the current data state (obtaining abnormal data) based on the AIGC capabilities of the large language model. It then outputs attribution hypotheses (i.e., candidate root causes) based on the model's reasoning capabilities. The output attribution hypotheses and current data state are then fed into the large language model, which then performs attribution analysis, normalizes the attribution conclusions, and ultimately outputs the attribution analysis results. The attribution analysis results output by the large language model can be cached as user history records or small samples to improve response efficiency.

[0050] This application embodiment provides an indicator change attribution analysis method, see Figure 2 ,by Figure 1 Taking the system shown as an example to execute the embodiment of the present application, the method flow provided in the embodiment of the present application includes:

[0051] 201. In response to an attribution analysis request for a target indicator change, query target data related to the target indicator.

[0052] The target indicator is a measurement parameter, such as GMV. An abnormality refers to a change or variation in an indicator, generally a change in the "dimension" level of the "indicator." An abnormality also has an implicit logic: "comparison." Generally speaking, an abnormality can be a comparison between the current period and the base period for the same entity, or a comparison between different entities over the same period. For example, if this month's GMV decreases compared to last month, this is a comparison between the current period and the base period for the same entity; if the conversion rate of the new version of the recommendation algorithm decreases compared to the old version, this is a comparison between different entities and over the same period.

[0053] When receiving an alert about a target indicator change, or when a user wants to find the root cause of the change, they can enter an attribution analysis request for the target indicator change on the interactive page. This request includes the target indicator, the current time for analysis, and the comparison time. For example, an attribution analysis request might be, "Help me attribute the GMV from February 5, 2025, to February 26, 2025, with the comparison time being February 17, 2024, to March 9, 2024."

[0054] Specifically, in response to an attribution analysis request for a target indicator change, querying target data related to the target indicator includes the following steps:

[0055] 2011. In response to an attribution analysis request for a target indicator change, identifying that the query intent for the target indicator is an attribution analysis intent.

[0056] Among them, the query intents supported in the embodiments of this application include attribution analysis, prediction, table search / number query, follow-up question, early warning, business report, etc. Specifically, in response to an attribution analysis request for a target indicator change, identifying the query intent for the target indicator as an attribution analysis intent may include:

[0057] The first step is to obtain historical prompt information.

[0058] In the embodiment of the present application, the user's historical prompt information is stored, and based on the historical prompt information, the user's intention can be better understood. The historical prompt information includes historical query statements and corresponding query intentions.

[0059] In the second step, the historical prompt information and attribution analysis request are input into the intent recognition model, and the query intent of the target indicator is output as the attribution analysis intent.

[0060] After obtaining the user's historical prompt information, the user's historical prompt information and query information can be input into the intent recognition model. The intent recognition model can identify the query intent of the target indicator as the attribution analysis intention by learning the user's historical query statements and their corresponding query intents.

[0061] 2012. Based on the attribution analysis intent and attribution analysis request, generate the target query statement corresponding to the target indicator.

[0062] Specifically, based on the attribution analysis intent and attribution analysis request, a target query statement corresponding to the target indicator is generated, including the following steps:

[0063] The first step is to determine the core parameters based on the attribution analysis request.

[0064] Core parameters refer to parameters relevant to attribution analysis of target indicators with changes. This step identifies keywords in the attribution analysis request and extracts core parameters from the identified keywords. For example, if the attribution analysis request is "Help me attribute GMV from 2025-02-05 to 2025-02-26, with a comparison period of 2024-02-17 to 2024-03-09," keyword identification can extract the core parameters "Current time: 2025-02-05 to 2025-02-26," "GMV," and "Comparison period: 2024-02-17 to 2024-03-09."

[0065] Optionally, if the attribution analysis request input by the user does not include a comparison time, or the current time included is unclear terms such as "recently" or "recently", a time-related prompt template can also be provided to the user for the user to select the comparison time and the specific time of the current time. From a practical perspective, the current time window and comparison time window corresponding to the attribution analysis request will cover working days and non-working days, and may also involve promotional activities and holidays, etc. The data related to the target indicators may include fake order data and non-fake order data, etc. For these response factors, corresponding prompt templates can be provided to the user, and the user can choose whether to consider these influencing factors when obtaining the target data. For example, the multiple prompt templates and corresponding options provided to the user can be:

[0066] 1. Do the selected time and comparison time need to be eliminated?

[0067] Eliminate major promotions and holidays. No special processing is required for holidays and major promotions.

[0068] 2. Do the selected time and comparison time require special processing?

[0069] View only working days View only non-working days No special processing is required to determine whether it is a working day or not

[0070] 3. Do the selected indicators require special processing?

[0071] Eliminate fake orders without special processing

[0072] When detecting that the user has selected the "No special handling for holidays and promotions" option in prompt template 1, "No special handling for working days" in prompt template 2, and "Eliminate fake orders" in prompt template 3, the resulting attribution analysis request may be "Help me attribute the GMV from 2025-02-05 to 2025-02-26, with the comparison period being 2024-02-17 to 2024-03-09. No special handling for working days, no special handling for holidays and promotions, and no elimination of fake orders."

[0073] Furthermore, for different options selected by the user, parameters can be extracted from the selected options and used as core parameters. For example, if the user selects "No special handling of holidays and promotions," "No special handling of weekdays," and "Exclude fake order data," then "No special handling of holidays and promotions," "No special handling of weekdays," and "Exclude fake order data" will be used as core parameters.

[0074] The second step is to determine the table name and fields of the target query statement based on the core parameters and attribution analysis intent.

[0075] In this step, you can use core parameters as fields and determine the table name based on the attribution analysis intent.

[0076] The third step is to generate the target query statement based on the table name and fields.

[0077] According to the grammatical structure corresponding to the query statement, the table name and field are filled in the position indicated by the grammatical structure to generate the target query statement.

[0078] 2013. Based on the target query statement, query the target data related to the target indicator.

[0079] Based on the generated target query statement and search engine, target data related to the target indicator can be queried from the database.

[0080] Furthermore, for the target data related to the target indicator, the target indicator in the current time window can be analyzed and compared with the target indicator in the previous time window to find problems with the target indicator in the current time window. For example, still using the above attribution analysis request as an example, the process of finding problems includes:

[0081] Based on your question, the current time window is [February 5, 2025 - February 26, 2025], which lasts 22 days. The comparison time window is [February 17, 2024 - March 9, 2024], which lasts 21 days. The resulting metric is GMV. Since the user requested to exclude fake orders, this data will be removed from the analysis. No special processing is required for holidays and major sales events.

[0082] "[Issue Found] Confirm Current Data Performance" is completed, and the query results are as follows:

[0083] During the abnormal movement detection, it was found that there was a significant abnormal movement between the analysis time and the comparison time you selected, and the indicator was in an abnormal decline range.

[0084] The actual value of the analysis period: the cumulative GMV value is 41175316147.96344 (41.175 billion), and the average value is 1871605279.4528837 (1.872 billion).

[0085] Compare the actual values ​​during the time period: The cumulative GMV value is 49481991632.2211 (49.482 billion), and the average is 2356285315.820052 (2.356 billion).

[0086] Forecast value for the analysis period: GMV average is 2.051 billion

[0087] Overall, the analysis period was 179 million lower than the forecast, a difference of -9.59%, exceeding the 5% significance level, indicating an abnormal decline. Furthermore, compared to the comparison period, the difference was -20.57%. Next, we will conduct attribution analysis on multiple topics, including buyers, sellers, orders, terminals, and product supply.

[0088] 202. Call multiple intelligent agents to analyze the target data from multiple dimensions to obtain abnormal data in multiple dimensions.

[0089] The multiple agents include at least one general agent and at least one vertical domain agent. Each agent can be mapped to a topic, and each topic is mapped to at least one agent. Topics can be understood as analysis angles, including buyers, sellers, orders, terminals, product supply, etc. Specifically, calling multiple agents to analyze target data from multiple dimensions to obtain abnormal data in multiple dimensions includes the following steps:

[0090] 2021. For any intelligent agent, call the intelligent agent and break down the target indicators into multiple dimensions.

[0091] For any agent, when calling the agent, you can use the indicator decomposition model to decompose the target indicator into multiple dimensions under the topic corresponding to the agent. Each dimension may have an impact on the change of the target indicator. For example, if the target indicator is GMV, under the buyer topic, GMV can be decomposed into multiple dimensions such as arpu (average receivables per user) level, activity level, transaction level, whether it is a new buyer, first call channel, and user bc (corporate user); under the buyer topic, GMV can be decomposed into multiple dimensions such as merchant years, merchant origin province, etc.; under the order topic, GMV can be decomposed into multiple dimensions such as order source; under the terminal topic, GMV can be decomposed into multiple dimensions such as app terminal; under the product supply topic, GMV can be decomposed into multiple dimensions such as category grouping, price power score, and category category.

[0092] 2022. Based on the target data corresponding to the multiple dimensions, the contribution of the multiple dimensions is calculated respectively.

[0093] The contribution of any dimension is used to represent the degree of contribution of the dimension to the change of the target indicator. When calculating the contribution of multiple dimensions based on the target data corresponding to the multiple dimensions, the existing contribution algorithm can be used for calculation.

[0094] 2023. From the contribution degrees of multiple dimensions, select at least one dimension whose contribution degree is greater than a preset contribution degree, and use the target data corresponding to the at least one dimension as abnormal data under the dimension.

[0095] Among them, the preset contribution degree can be set by technical personnel.

[0096] 203. Input abnormal data from multiple dimensions into the attribution reasoning model and output multiple candidate root causes that lead to changes in target indicators.

[0097] The attribution reasoning model is used to reason based on abnormal data across multiple dimensions, identifying multiple candidate root causes for indicator fluctuations. Due to the powerful reasoning capabilities of the large language model, the attribution reasoning model can reason from multiple thematic perspectives and even conduct divergent reasoning.

[0098] Still taking the above example, the specific reasoning process of the attribution reasoning model is:

[0099] The results of general thematic reasoning from the buyer’s perspective are as follows:

[0100] Because this is a thematic analysis based on the buyer's perspective, we first conducted a dimensional analysis of the buyer-perspective labels: ARPU level, activity level, transaction level, new buyer status, first call channel, and user BC classification. We discovered the following significant contributing combinations: [Inactive Group, Low Activity Group] in the activity level, and [High ARPU Group] in the ARPU level. Furthermore, separate analysis was conducted on [Super Buyers, High-Small Buyers, and Professional Buyers] within the buyer category, revealing no significant anomalies among buyers in these specific dimensions. Based on the dimensional conclusions, we analyzed the indicators and found that the overall GMV decline was primarily driven by average order value, which was primarily driven by a decrease in average transaction value. Within the GMV contributed by inactive or low-activity buyers, the decrease in average transaction value was also significant.

[0101]

Conclusion

[0102] The decline in GMV contributed by buyers belonging to the inactive and low-activity groups in the activity level was the most affected, with their daily average value dropping from 1.253 billion to 849 million, with a contribution rate of 72.85% and an excess contribution rate of 17.22%.

[0103] The GMV contributed by buyers in the high ARPU group in the ARPU level also had a significant impact. Its average daily value dropped from 1.437 billion to 1.076 billion, with a contribution rate of 65.16%, but the excess contribution rate was only 1.33%, indicating that there was no significant excess impact.

[0104] A separate analysis of super buyers, high-end buyers, and professional buyers who require specific domain labels revealed the following:

[0105] There is no significant difference in contribution between super buyers and non-super buyers, indicating that the super buyer group is relatively stable.

[0106] There is no significant difference in contribution between high-school-educated buyers and non-high-school-educated buyers, indicating that the high-school-educated buyer group is relatively stable.

[0107] Compared with non-professional buyers, professional buyers are less affected by GMV fluctuations and have lower contributions.

[0108] We analyzed the buyer-side indicators for both the overall and high-contribution GMV and found that:

[0109] Overall, the decline in GMV was primarily driven by average order value, which contributed 133% to the decline, from 1,425.41 to 1,049.03. Further analysis revealed that average order value was the primary contributing factor, contributing 113% to the decline, from 171.08 to 131.73.

[0110] Within the GMV contributed by inactive or low-activity buyers, average order value also significantly impacted the decline, contributing 120%, dropping from 1,401.05 to 937.56. Further analysis revealed that average order value was the root cause, contributing 80%, dropping from 213.35 to 163.31.

[0111] The results of general topic reasoning from the seller’s perspective are as follows:

[0112] Because this is a thematic analysis based on the seller's perspective, we first conducted a dimensional flattening analysis of the two seller-perspective labels, Merchant Year and Merchant Province of Origin. We identified the following notable combinations of significant contributions: [Unknown] in Merchant Year, and Merchant Province of Origin * [Guangdong] * [Unknown] in Merchant Year. Because Merchant Year and Merchant Province of Origin lack further sub-dimensions, no drill-down analysis was required. Furthermore, separate analysis was required for the sellers' categories of [Specific Channel Merchants, Powerful Merchants, and Super Factory Merchants]. We found that merchants in these special dimensions did not contribute significantly, while merchants outside these special dimensions actually contributed more. We then conducted an indicator breakdown based on the dimensional conclusions, but since the indicator breakdown conclusions were empty, further analysis was not possible.

[0113]

Conclusion

[0114] The decline in GMV contributed by merchants with unknown merchant years was the most affected, with their daily average value dropping from 2.356 billion to 553,200, with a contribution rate of 486.04% and an excess contribution rate of 386.04%.

[0115] Merchants whose origin province is Guangdong and whose business years are unknown also contribute a lot to GMV. Their daily average value dropped from 861 million to 77,200, with a contribution rate of 177.72% and an excess contribution rate of 141.16%.

[0116] A separate analysis of sellers in specific channels, powerful merchants, and super factory merchants who require specific field labels revealed the following:

[0117] Both specific channel merchants and non-specific channel merchants have seen a significant decline, which is not an influencing factor.

[0118] Compared with non-strong merchants, strong merchants are less affected by GMV fluctuations, and non-strong merchants contribute more.

[0119] Super factory merchants are less affected by GMV fluctuations than non-super factory merchants, and non-super factory merchants contribute more.

[0120] The GMV of the overall and high contribution dimension value is decomposed into seller-side indicators, but the index decomposition conclusion is empty, and further analysis cannot be performed.

[0121] The perspective general topic reasoning result is as follows:

[0122] Because it is an order perspective topic analysis, first of all, the order source dimension is analyzed, and the following several combinations of significant contribution items that are worth attention are found:

others

others

[0123]

Conclusion

[0124] The order source belonging to others contributes the most to the decline of GMV, with a daily average of 1.477 billion, a contribution degree of 129.74%, and an excess contribution degree of 41.61%.

[0125] The analysis of commercial orders, distribution orders, factory finding orders, cross-border GMV, and contract orders finds:

[0126] The contribution degree of these special domain label orders is low, and the contribution degree of orders outside these special fields is high, indicating that these special fields are relatively stable.

[0127] Among them, the order quantity of commercial orders, distribution orders, factory finding orders, cross-border GMV, and contract orders does not show a significant decline, and the performance is relatively stable.

[0128] The index decomposition of the overall GMV finds:

[0129] The decline of the overall GMV is mainly affected by the pen price, which decreases from 171.08 to 131.73, with a contribution degree of 113%, indicating that the pen price plays a major role in the decline of the overall GMV.

[0130] And the GMV decomposition of order source

others

[0131] The results of general topic reasoning from the end perspective are as follows:

[0132] Because this is a thematic analysis based on an end-user perspective, we first conducted a flat analysis of the end-user type dimension. We discovered the following significant contributing item combination: [app] within the end-user type dimension. Since the end-user type dimension has no corresponding secondary dimension, no drill-down analysis is required. The Special Dimension Conclusion section is empty, indicating that no special dimensions require separate analysis. The Indicator Breakdown Conclusion section is empty, indicating that no further indicator breakdown analysis results are available.

[0133] [Conclusion] An analysis of various dimensions related to the client side revealed the following:

[0134] The decline in GMV on the app side had the greatest impact, with its daily average value dropping from 1.227 billion to 940 million, with a contribution rate of 59.1% and an excess contribution rate of 7.05%.

[0135] Because the end type dimension has no secondary dimensions, no further analysis is required.

[0136] A separate analysis of specific dimensions revealed:

[0137] No special dimensions were found that required special attention.

[0138] By breaking down the overall GMV into device-side indicators, we found the following:

[0139] Since the indicator decomposition conclusion is empty, no significant contributing factors were found.

[0140] The results of general thematic reasoning from the commodity supply perspective are as follows:

[0141] Because this is a thematic analysis based on the product supply perspective, we first conducted a dimensional analysis of the product supply-based labels—category grouping, price power score, and category broad categories. We identified the following significant contributing combinations: the price power score combination of "no price power, extremely high price power" and the category broad category "consumer goods." Since no secondary dimensions were identified for further drill-down, no secondary dimension analysis was conducted. Regarding the special dimensions of whether a product is commercialized, whether it is a gold crown product, whether it is a potential product, whether it is a hyperlink product, whether it is customized, whether it is an expert-selected product, and whether it is a carefully selected product, we found that these special dimensions had low contributions, while non-special dimensions had high contributions. Therefore, these special dimensions are relatively robust and do not require special attention. Based on the dimensional conclusions, we analyzed the indicators and found that the overall GMV decline was primarily influenced by the number of actively selling products. Further analysis revealed that the active sales rate was the primary root cause of the GMV decline. Furthermore, the GMV decline of consumer goods products with no price power or extremely high price power was also primarily influenced by the number of actively selling products. Further analysis revealed that the active sales rate was the primary root cause of the GMV decline.

[0142]

Conclusion

[0143] Among the price power scores, those with no price power, those with extremely high price power, and whose major category is consumer goods have the greatest impact on the decline in GMV, with their daily average value dropping from 371 million to 149 million, with a contribution rate of 45.7% and an excess contribution rate of 29.96%.

[0144] Analysis of special dimensions revealed:

[0145] Special dimensions such as commercial products, gold crown products, potential products, hyperlink products, Wow customized products, expert selected products, and strictly selected products are relatively stable and do not require special attention.

[0146] By breaking down the overall GMV indicators, we found that:

[0147] In the first level of analysis, the contribution of the number of moving goods indicator was 75%, which dropped from 4.4754 million to 3.7619 million, indicating that the number of moving goods had a greater impact on the overall GMV.

[0148] At the second level of analysis, the active sales rate, a key contributor to the current path, contributed 122%, a decrease from 22% to 17%. Since there are no further paths to break down, the active sales rate can be considered the primary indicator of GMV fluctuations.

[0149] By analyzing the GMV contributed by consumer goods with no or very high price power, we found that:

[0150] In the first level of analysis, the contribution of the number of moving goods indicator was 96%, which dropped from 1.1192 million to 469,600, indicating that the number of moving goods had a certain impact on the overall GMV decline.

[0151] At the second level of analysis, the active sales rate, a key contributor to the current path, contributed 81%, a decrease from 18% to 9%. Since there are no further paths to break down, the active sales rate can be considered the primary indicator of the GMV decline.

[0152] 204. Input multiple candidate root causes, abnormal data in multiple dimensions, and data from a preset knowledge base into the attribution analysis model, and output the attribution analysis results of the target indicator changes.

[0153] The preset knowledge base is used to store descriptive data for different indicators and dimensions, such as the definitions of "carefully selected products," "newcomers," and "old customers." It can also store domain-related data, such as promotional events held during a certain period. The preset knowledge base can be divided into a domain knowledge base and a professional knowledge base. The domain knowledge base is used to store domain-related data, while the professional knowledge base stores descriptive data for different indicators and dimensions. The data stored in the knowledge base can be dynamically updated based on actual scenarios. When multiple candidate root causes, abnormal data across multiple dimensions, and data from the preset knowledge base are input into the attribution analysis model, the attribution analysis model leverages the powerful reasoning capabilities of the large language model to infer the root causes that may have caused the indicator fluctuations based on these multiple candidate root causes, abnormal data across multiple dimensions, and data from the preset knowledge base, and derive the root causes that may have caused the indicator fluctuations. The model then normalizes these root causes based on the AIGC function of the large language model, resulting in the attribution analysis results.

[0154] Attribution analysis results for target indicator changes based on the attribution analysis model include first-category attribution analysis results, second-category attribution analysis results, and third-category attribution analysis results. First-category attribution analysis results refer to those verified using the verification methods provided by the attribution analysis model. Second-category attribution analysis results refer to those not verified using the verification methods provided by the attribution analysis model. Third-category attribution analysis results refer to those for which the attribution analysis model does not provide a verification method. When an attribution analysis model provides a verification method for attribution analysis results, and the corresponding data is obtained from the database according to the verification method, and the attribution analysis results given by the attribution analysis model are verified based on the obtained data, it is indicated that the attribution analysis results are relatively reliable, and the attribution analysis results given by the attribution analysis model belong to the first category of attribution analysis results. When an attribution analysis model provides a verification method for attribution analysis results, but the corresponding data cannot be obtained from the database according to the verification method, and the attribution analysis results given by the attribution analysis model cannot be verified based on the obtained data, it is indicated that the attribution analysis results are unreliable, and the attribution analysis results given by the attribution analysis model belong to the second category of attribution analysis results. When the attribution analysis model does not provide a verification method for attribution analysis results and the attribution analysis results cannot be verified, it is indicated that the attribution analysis results are unreliable, and the attribution analysis results given by the attribution analysis model belong to the third category of attribution analysis results.

[0155] Furthermore, for the second type of attribution analysis results, the attribution analysis model can also be fed back to the intelligent agent to decompose the target indicators into dimensions used to verify the second type of attribution analysis results, and increase the data of this dimension, thereby realizing the verification of the second type of attribution analysis results, making the second type of attribution analysis results become the first type of attribution analysis results, and improving the reliability of the attribution analysis results.

[0156] Still taking the above example, the reasoning and verification process of the attribution analysis model includes:

[0157] In response to a user's question about attributing GMV fluctuations after removing fake orders from February 5 to 26, 2025 (the comparison period is February 17 to March 9, 2024), we proposed and verified several hypotheses. The specific scenarios are as follows:

[0158]

Hypothesis 1

[0159] Rationale: The sharp drop in orders from [Other] sources on the order side (daily average GMV dropped from 2.077 billion to 1.448 billion) resonated with a significant decline in app-based GMV on the buyer side (contribution: 59.1%). This may be related to the digital marketing investment strategy implemented in February 2025. This strategy emphasizes paid traffic and targeted targeting, which may have led to a decrease in the proportion of organic traffic sources (such as homepage recommendations). (Special Feature Hypothesis)

[0160] [Verification direction]: It is necessary to verify the changes in the proportion of natural traffic and the conversion efficiency data of paid traffic during the same period.

[0161] [Conclusion]: This hypothesis cannot be proven at present, and additional data is needed on the changes in the proportion of natural traffic and the conversion efficiency of paid traffic.

[0162] Hypothesis 2: Adjustments in price power rules affect the exposure of consumer goods

[0163] [Argument 2]: The supply side of the product shows a sharp drop in the sales rate of consumer products with no or very high price power (from 22% to 17%), which is correlated with a decrease in the unit price of orders (171.08 → 131.73). It's possible that the platform adjusted its price power calculation rules in early 2025, causing some products to lose search weight. The specific details of this adjustment need to be checked in the internal policy library. (Special Feature Hypothesis)

[0164] [Verification direction]: It is necessary to check whether any price power-related strategy announcements were issued in January-February 2025.

[0165]

Conclusion

[0166] [Hypothesis 3]: Changes to the new merchant onboarding process lead to qualification certification anomalies

[0167] [Basis]: The GMV of merchants whose business years are displayed as "Unknown" on the seller side has plummeted (from 861 million to 77,200 in Guangdong). Combined with the progress of the Merchant Development Center activities reported in the knowledge base, there may be new regulations forcing merchants to complete qualification information such as business years, and merchants who have not completed certification may be subject to traffic restrictions. (Special Feature Hypothesis)

[0168] [Verification direction]: It is necessary to verify the change records of merchant qualification review strategy in Q1 2025.

[0169] [Conclusion]: It is assumed that "changes in the new merchant entry process lead to abnormal qualification certification" can be proved.

[0170] Hypothesis 4: Macroeconomic downward pressure leads to procurement delays for SMEs

[0171]

Basis

[0172] [Verification direction]: It is necessary to verify whether the GMV decline of the platform's core industrial product categories (such as steel and plastic raw materials) during the same period is higher than the market average.

[0173]

Conclusion

[0174]

Hypothesis

[0175] [Basis]: A search revealed that aluminum tariffs will increase in February 2025. Cross-border buyers on 1688 (who contribute approximately 30%) may reduce their raw material purchases due to export restrictions. This requires verification of the GMV decline of cross-border buyers during the same period and related product category fluctuations. (Divergent hypothesis)

[0176]

Verification direction

[0177] Conclusion: The hypothesis that intensified trade frictions may be affecting the stability of cross-border supply chains cannot be proven. The thematic analysis shows that cross-border GMV in specific sectors remained stable, with no significant decline. Furthermore, the platform held several events promoting cross-border commerce in February 2025, indicating increased support for cross-border commerce during this period. Additional information is needed: GMV declines for specific cross-border buyers and fluctuations in related categories, as well as more detailed information on the actual impact of trade frictions on the platform's cross-border commerce.

[0178] After normalizing the root causes obtained above, the final attribution analysis results are as follows:

[0179] The user's question is analyzed from multiple perspectives such as the user, order, merchant, and commodity supply, and further hypotheses are proposed and verified. Finally, it is found that:

[0180] 1. Price force rule adjustment affects consumer goods exposure. The dynamic sales rate of consumer goods without price force / very high price force drops from 22% to 17%. This is related to the decrease in order-side unit price (171.08→131.73). This may be due to the platform adjusting the price force calculation rules in early 2025, resulting in some goods losing search weighting. Through internal strategy library queries, it is found that price force calculation rules were indeed released in January-February 2025, resulting in some goods losing search weighting. Through internal strategy library queries, it is found that price force-related strategy announcements were indeed released in January-February 2025, which may directly lead to a decrease in commodity exposure, thereby affecting overall GMV.

[0181] 2. Changes in the process of new merchants joining the platform result in abnormal qualification certification. The GMV of the group of sellers showing the business age 'unknown' drops significantly (from 861 million to 772 million in Guangdong). This may be due to changes in merchant qualification audit strategies in Q1 2025, which require merchants to complete qualification information such as business age. Merchants who have not completed the certification are restricted from traffic. After checking the records of merchant qualification audit strategy changes in Q1 2025, it is found that there are indeed related adjustments, which may directly affect the traffic distribution of some merchants, thereby leading to a decrease in GMV.

[0182] 3. The GMV contributed by the inactive and low-active groups in the active level has decreased significantly (daily average from 12.53 billion to 8.49 billion, contribution 72.85%). At the same time, the overall GMV decline is mainly affected by the average order value, which has decreased from 171.08 to 131.73. This may be related to the adjustment of platform traffic distribution strategies. Although this hypothesis cannot be directly proven at present, combined with buyer-side data, it can be seen that the loss of low-active buyers and the decrease in unit price have a significant impact on GMV.

[0183] 4. In addition, app-side GMV has decreased significantly (daily average from 12.27 billion to 9.4 billion, contribution 59.1%). This may be related to the reduction in natural traffic entry due to platform traffic distribution strategy adjustments, but currently lacks direct data support and needs further verification of natural traffic proportion changes and paid traffic conversion efficiency data.

[0184] In addition to the above topic-based reasoning, multiple possible influencing factors are also considered for reference: a. Based on the analysis of the retrieved manufacturing PMI, it is speculated that the downward pressure on the macro economy may lead to delayed procurement by small and medium-sized enterprises, and it needs to be verified whether the GMV of the platform's core industrial product categories decreased more than the average of the overall market during the same period. b. Based on the retrieved aluminum tariff policy information, it is speculated that the intensification of trade friction may affect the stability of cross-border supply chains, and it needs to be verified whether the GMV of cross-border buyers and the fluctuation data of related categories decreased during the same period.

[0185] In addition to giving the attribution analysis results of the target indicator anomaly, the attribution analysis model can also give corresponding guidance suggestions.

[0186] For example, the suggestions that can be given for the above example are:

[0187] 1. For the impact of price force rule adjustment, the platform is suggested to optimize the price force calculation rules to ensure that rule adjustments do not have a significant impact on product exposure. At the same time, strengthen the training of merchants on price force, help merchants better adapt to the new rules.

[0188] 2. For the qualification certification abnormal problem caused by the change of the new merchant registration process, the platform is suggested to optimize the merchant qualification audit process to reduce the traffic restrictions caused by incomplete qualification information. At the same time, strengthen the support and service for new merchants, help them quickly adapt to the platform rules.

[0189] 3. For the problem of low active buyer loss and single price decline, the platform is suggested to strengthen marketing activities for low active buyers, such as launching exclusive coupons or member benefits, to improve their activity and purchasing power. At the same time, optimize the product structure, provide more cost-effective goods, and attract buyers to increase the average order value.

[0190] 4. Continue to pay attention to changes in macroeconomic indicators and adjust platform strategies in a timely manner to cope with the impact of small and medium-sized enterprise procurement budget tightening. For cross-border business, it is suggested to strengthen communication with overseas buyers, understand their procurement needs and pain points, optimize product supply and fulfillment services, and improve the cross-border buyer's procurement experience.

[0191] Further, after inputting multiple candidate root causes, abnormal data in multiple dimensions, and data in the preset knowledge base into the attribution analysis model, outputting the attribution analysis results of the target indicator anomaly, and based on the front-end service capability, a visual card can be built to display the attribution analysis results of the target indicator in the visual card.

[0192] Further, after the attribution analysis results of the target indicator anomaly are displayed, the attribution analysis results of the target indicator anomaly can be cached, so that when a user wants to query the attribution analysis results of the target indicator anomaly in the future, the results can be directly provided to the user.

[0193] Figure 3 A flowchart of the target indicator change attribution analysis method provided in an embodiment of the present application is shown. Figure 3 , the overall analysis process includes:

[0194] The system obtains a user's query for a specific metric and, based on historically enhanced prompts (i.e., historical prompt information), uses an intent recognition model to identify the user's query intent. If the query intent is identified as attribution analysis, it uses prompt engineering to identify core parameters based on the query input and common sense about time and other dimensions. Based on the identified intent and core parameters (including time, metric, and dimension), it infers the fields and table names of the SQL statement, generating a SQL statement. Based on the SQL statement, it uses a data search engine to retrieve target data related to the metric. Multiple agents analyze the target data, generating anomaly data across multiple dimensions. A large language model is then used to infer these anomaly data, generating multiple hypotheses (candidate root causes) that could lead to the metric fluctuation. These hypotheses, the anomaly data across multiple dimensions, and data from the professional knowledge base and domain knowledge base are then fed into the large language model. Leveraging the AIGC capabilities of the large language model, it outputs the attribution analysis results for the metric fluctuation. A visualization card is then constructed to display these attribution analysis results. Furthermore, the attribution analysis results are cached so that they can be directly provided when users make subsequent inquiries, thereby improving response efficiency.

[0195] All of the above optional technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here.

[0196] Please refer to Figure 4 , which shows a schematic diagram of the structure of an indicator change attribution analysis device provided in an embodiment of the present application. The device can be implemented through software, hardware, or a combination of both, and becomes all or part of an electronic device. The device includes:

[0197] A query module 401 is configured to query target data related to a target indicator in response to an attribution analysis request for a target indicator change;

[0198] An analysis module 402 is configured to call multiple agents to analyze the target data from multiple dimensions to obtain abnormal data in multiple dimensions;

[0199] The first input-output module 403 is configured to input the abnormal data in the multiple dimensions into the attribution reasoning model and output multiple candidate root causes that lead to the abnormal change of the target indicator;

[0200] The second input-output module 404 is used to input the multiple candidate root causes, the abnormal data under the multiple dimensions, and the data of the preset knowledge base into the attribution analysis model, and output the attribution analysis results of the target indicators. The preset knowledge base is used to store descriptive data and field-related data of different indicators and dimensions.

[0201] In another embodiment of the present application, the query module 401 is used to respond to an attribution analysis request for a change in a target indicator, identify that the query intention for the target indicator is an attribution analysis intention; generate a target query statement corresponding to the target indicator based on the attribution analysis intention and the attribution analysis request; and query target data related to the target indicator based on the target query statement.

[0202] In another embodiment of the present application, the query module 401 is used to obtain historical prompt information, which includes historical query statements and corresponding query intentions; input the historical prompt information and the attribution analysis request into the intent recognition model, and output the query intention of the target indicator as the attribution analysis intention.

[0203] In another embodiment of the present application, the query module 401 is used to determine core parameters based on the attribution analysis request, where the core parameters refer to parameters related to the attribution analysis of the target indicators that have changed; determine the table name and fields of the target query statement based on the core parameters and the attribution analysis intention; and generate the target query statement based on the table name and fields.

[0204] In another embodiment of the present application, the calling module 402 is used to call any intelligent agent and decompose the target indicator into multiple dimensions; based on the target data corresponding to the multiple dimensions, the contribution of the multiple dimensions is calculated respectively, and the contribution of any dimension is used to characterize the degree of contribution of the dimension to the change of the target indicator; from the contribution of the multiple dimensions, at least one dimension with a contribution greater than a preset contribution is selected, and the target data corresponding to the at least one dimension is used as the abnormal data under the dimension.

[0205] In another embodiment of the present application, the attribution analysis results of the target indicator include a first-category attribution analysis result, a second-category attribution analysis result, and a third-category attribution analysis result:

[0206] The first type of attribution analysis results refers to attribution analysis results that have been verified by the verification method provided by the attribution analysis model;

[0207] The second type of attribution analysis results refers to attribution analysis results that have not been verified by the verification method provided by the attribution analysis model;

[0208] The third type of attribution analysis results refers to attribution analysis results for which the attribution analysis model does not provide a verification method.

[0209] In another embodiment of the present application, the device further comprises:

[0210] The display module is used to display the attribution analysis results of the target indicators.

[0211] Figure 5 FIG2 shows a block diagram of an electronic device 500 according to an exemplary embodiment of the present application. Generally, the electronic device 500 includes a processor 501 and a memory 502 .

[0212] The processor 501 can be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 501 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state; the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 501 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 501 may also include an artificial intelligence processor, which is used to process computing operations related to machine learning.

[0213] The memory 502 may include one or more computer-readable storage media, which may be non-transitory computer-readable storage media, such as CD-ROMs (Compact Disc Read-Only Memory), ROMs, RAMs (Random Access Memory), magnetic tapes, floppy disks, and optical data storage devices. The computer-readable storage media may store at least one computer program, which, when executed, can implement the aforementioned indicator change attribution analysis method.

[0214] Of course, the electronic device described above may also include other components, such as input / output interfaces and communication components. The input / output interface provides an interface between the processor and a peripheral interface module, which may be an output device, an input device, etc. The communication component is configured to facilitate wired or wireless communication between the electronic device and other devices.

[0215] Those skilled in the art will understand that Figure 5 The structure shown in the figure does not constitute a limitation on the electronic device 500, and the electronic device 500 may include more or fewer components than shown in the figure, or combine certain components, or adopt a different component arrangement.

[0216] An embodiment of the present application provides a computer-readable storage medium, in which at least one computer program is stored. When the at least one computer program is executed by a processor, the above-mentioned indicator change attribution analysis method can be implemented.

[0217] An embodiment of the present application provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it can implement the above-mentioned indicator change attribution analysis method.

[0218] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0219] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for attribution analysis of indicator changes, characterized in that: The method comprises: In response to an attribution analysis request for a target indicator change, querying target data related to the target indicator; Calling multiple intelligent agents to analyze the target data from multiple dimensions to obtain abnormal data in multiple dimensions; Inputting the abnormal data under the multiple dimensions into the attribution reasoning model to output multiple candidate root causes that lead to the abnormal changes in the target indicators; The multiple candidate root causes, the abnormal data under the multiple dimensions and the data of the preset knowledge base are input into the attribution analysis model, and the attribution analysis results of the target indicator changes are output. The preset knowledge base is used to store descriptive data and field-related data of different indicators and dimensions.

2. The method according to claim 1, characterized in that The step of querying target data related to the target indicator in response to the attribution analysis request for the target indicator change includes: In response to the attribution analysis request for the target indicator change, identifying that the query intent for the target indicator is an attribution analysis intent; Based on the attribution analysis intention and the attribution analysis request, generating a target query statement corresponding to the target indicator; Based on the target query statement, target data related to the target indicator is queried.

3. The method according to claim 2, characterized in that The step of responding to the attribution analysis request for the target indicator change and identifying that the query intent for the target indicator is an attribution analysis intent includes: Obtain historical prompt information, wherein the historical prompt information includes historical query statements and corresponding query intents; The historical prompt information and the attribution analysis request are input into an intent recognition model, and the query intent of the target indicator is output as the attribution analysis intent.

4. The method according to claim 2, characterized in that The generating, based on the attribution analysis intention and the attribution analysis request, a target query statement corresponding to the target indicator includes: Determining core parameters based on the attribution analysis request, where the core parameters refer to parameters related to attribution analysis of the target indicator of the change; Determine the table name and field of the target query statement based on the core parameters and the attribution analysis intent; The target query statement is generated based on the table name and fields.

5. The method according to claim 1, wherein The calling of multiple agents to analyze the target data from multiple dimensions to obtain abnormal data in multiple dimensions includes: For any intelligent agent, the intelligent agent is called to decompose the target indicator into multiple dimensions; Based on the target data corresponding to the multiple dimensions, the contribution of the multiple dimensions is calculated respectively. The contribution of any dimension is used to represent the contribution degree of the dimension to the change of the target indicator; From the contribution degrees of the multiple dimensions, at least one dimension whose contribution degree is greater than a preset contribution degree is selected, and target data corresponding to the at least one dimension is used as abnormal data under the dimension.

6. The method according to claim 1, characterized in that The attribution analysis results of the target indicators include the first category attribution analysis results, the second category attribution analysis results and the third category attribution analysis results: The first type of attribution analysis results refers to attribution analysis results that have been verified by the verification method provided by the attribution analysis model; The second type of attribution analysis results refers to attribution analysis results that have not been verified by the verification method provided by the attribution analysis model; The third type of attribution analysis results refers to attribution analysis results for which the attribution analysis model does not provide a verification method.

7. The method according to any one of claims 1 to 6, characterized in that After inputting the multiple candidate root causes, the abnormal data in the multiple dimensions, and the data in the preset knowledge base into the attribution analysis model and outputting the attribution analysis result of the target indicator change, the method further includes: Displays the attribution analysis results of the target indicator.

8. An electronic device, characterized in that: It includes a processor and a memory; the memory stores at least one program code; the at least one program code is used to be called and executed by the processor to implement the indicator change attribution analysis method according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one computer program, and when the at least one computer program is executed by a processor, it can implement the indicator change attribution analysis method according to any one of claims 1 to 7.

10. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, it can implement the indicator change attribution analysis method according to any one of claims 1 to 7.

Images