Data analysis method and device and electronic equipment
By using a multi-agent architecture to process advertising data analysis and leveraging the combined invocation of multiple agents, the problem of low accuracy in complex tasks in traditional single LLMs is solved, achieving higher data analysis accuracy and task adaptability.
Patent Information
- Application Number
- CN202511243240.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-02
- Publication Date
- 2025-12-16
AI Technical Summary
Traditional single LLM models suffer from problems such as low data analysis accuracy, high error rate, difficulty in handling cross-domain knowledge, poor task adaptability, and difficulty in expansion in complex data analysis tasks.
A multi-agent architecture is adopted, which obtains analysis requirements by asking questions in natural language, selects multiple agents and determines their calling order, and processes advertising data in sequence to improve analysis accuracy.
It reduces the error rate of data retrieval, logical reasoning, and conclusion generation in a single LLM process, improves the accuracy of advertising data analysis, especially in complex tasks, reduces cross-domain knowledge comprehension bias and model illusion, and enhances task adaptability and the flexibility of tool expansion.
Smart Images
Figure CN121146836A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent analysis technology, and in particular to a data analysis method, apparatus, and electronic device. Background Technology
[0002] The field of data analytics is currently facing unprecedented pressure for transformation, with traditional analytical tools struggling to cope with the exponentially increasing complexity of data and the need for real-time decision-making. With the breakthrough development of Large Language Model (LLM) technology, intelligent agent systems built with LLM as their "central nervous system" are becoming the core direction of the next generation of data analytics infrastructure.
[0003] Current mainstream data analysis architectures primarily adopt the Single-agent model (such as ChatBI), whose basic workflow is: user inputs natural language query → LLM understands data structure → generates SQL / API calls → executes query → directly returns results. While this architecture performs reasonably well in simple query scenarios, it exposes a series of structural flaws in complex analysis tasks, such as broken complex reasoning chains: when a single LLM needs to handle multiple steps such as data retrieval, logical reasoning, and conclusion generation simultaneously, the error rate accumulates exponentially. In other words, the current single-agent data analysis architecture leads to low data analysis accuracy. Summary of the Invention
[0004] The purpose of this invention is to provide a data analysis method, apparatus, and electronic device to improve the accuracy of data analysis. The specific technical solution is as follows:
[0005] Firstly, a data analysis method is provided, including:
[0006] Acquire advertising data to be analyzed and a question expressed in natural language, wherein the question is used to express the analysis needs for the advertising data to be analyzed;
[0007] Determine the analysis type corresponding to the analysis requirement represented by the question;
[0008] Based on the functional description information of multiple agents, select multiple agents that match the analysis type and the calling order of each agent as the target configuration information;
[0009] According to the target configuration information, multiple agents are invoked sequentially to process the advertising data to be analyzed, and the data analysis results are obtained.
[0010] Optionally, determining the analysis type corresponding to the analysis requirement represented by the question; selecting the agent that matches the required analysis type and the calling order of each agent based on the pre-configured functional description information of multiple agents, as the target configuration information, includes:
[0011] The target analysis strategy type corresponding to the question is determined by the question classifier;
[0012] Find the correspondence between the configuration information corresponding to different analysis strategy types to obtain the target configuration information corresponding to the target analysis strategy type.
[0013] Optionally, before finding the correspondence between configuration information for different analysis strategy types, the method further includes:
[0014] For each analysis strategy type, identify the identifiers of multiple pre-configured agents;
[0015] The system receives the selection of identifiers for multiple pre-configured agents and determines the selection order of the agents, determines the invocation order of multiple agents corresponding to the analysis strategy type, and saves the combination and invocation order of multiple agents corresponding to the analysis strategy type.
[0016] Optionally, the multiple agents include at least two of the following: indicator group attribution agent, dimension group attribution agent, dimension drill-down agent, anomaly analysis agent, and customer business logic attribution agent;
[0017] The attribution agent is used to determine the core contribution metrics from the metrics that affect advertising revenue. The metrics that affect advertising revenue include: daily active users (DAU), average inventory per user, inventory fill rate, and advertising revenue (ecpm) per thousand impressions.
[0018] The dimension group attribution agent is used to identify the core influencing dimensions from those that affect advertising revenue;
[0019] The dimension drill-down agent is used to determine the fine-grained impact data of dimensions through the layer-by-layer decomposition of the dimension hierarchy structure;
[0020] The anomaly analysis agent is used to identify anomalies in the advertising data;
[0021] The customer business logic attribution agent is an attribution agent customized according to user needs.
[0022] Optionally, the target pre-configuration information includes the combination and invocation order of multiple intelligent agents corresponding to the analysis strategy used for advertising revenue analysis.
[0023] Secondly, a data analysis device is provided, comprising:
[0024] The acquisition module is used to acquire advertising data to be analyzed and questions expressed in natural language, wherein the questions are used to express the analysis needs for the advertising data to be analyzed;
[0025] The determination module is used to determine the analysis type corresponding to the analysis requirement represented by the question; and select multiple agents that match the analysis type and the calling order of each agent as target configuration information based on the functional description information of multiple agents.
[0026] The analysis module is used to sequentially call multiple agents according to the target configuration information to process the advertising data to be analyzed and obtain data analysis results.
[0027] Optionally, the determining module is specifically used to determine the target analysis strategy type corresponding to the question through a question classifier; find the correspondence between the configuration information corresponding to different analysis strategy types, and obtain the target configuration information corresponding to the target analysis strategy type.
[0028] Optionally, the device further includes:
[0029] The configuration module is used to determine the identifiers of multiple pre-configured agents for each analysis strategy type before searching for the correspondence of configuration information corresponding to different analysis strategy types; receive the selection of the identifiers of the multiple pre-configured agents and determine the selection order of the agents, determine the calling order of the multiple agents corresponding to the analysis strategy type, and save the combination and calling order of the multiple agents corresponding to the analysis strategy type.
[0030] Thirdly, an electronic device is provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;
[0031] Memory, used to store computer programs;
[0032] When a processor executes a program stored in memory, it implements any of the steps described in the first aspect.
[0033] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the steps of any of the methods described in the first aspect.
[0034] In another aspect of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and the computer program, when executed by a processor, implements any of the data analysis methods described above.
[0035] In another aspect of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the data analysis methods described above.
[0036] In this embodiment of the invention, by acquiring advertising data to be analyzed and a question expressed in natural language, the question represents the analysis requirement for the advertising data to be analyzed. The analysis type corresponding to the analysis requirement represented by the question is determined. Based on the functional description information of multiple agents, multiple agents matching the analysis type and the calling order of each agent are selected as target configuration information. This determines the target configuration information corresponding to the user's question expressed in natural language. The target configuration information includes the combination and calling order of multiple intelligent agents. Thus, multiple agents can be called sequentially according to the target configuration information to process the advertising data to be analyzed and obtain data analysis results. Each agent performs its own function. The combined action of multiple agents can reduce the exponentially accumulating error rate caused by a single LLM simultaneously handling data retrieval, logical reasoning, and conclusion generation, thereby improving the accuracy of advertising data analysis. This improvement in accuracy is particularly significant when facing complex tasks. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0038] Figure 1 This is a flowchart of the data analysis method in an embodiment of the present invention;
[0039] Figure 2 for Figure 1 A flowchart of S12 in China;
[0040] Figure 3 This is a flowchart illustrating the pre-setting of configuration information corresponding to each analysis strategy type in an embodiment of the present invention;
[0041] Figure 4 This is a schematic diagram illustrating the application of the data analysis method described in this embodiment of the invention.
[0042] Figure 5 This is a schematic diagram of the execution logic of the agent in an embodiment of the present invention;
[0043] Figure 6 This is a schematic diagram of the data analysis device in an embodiment of the present invention;
[0044] Figure 7 This is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. Detailed Implementation
[0045] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.
[0046] The technological evolution of intelligent agent systems built around LLM as the "central nervous system" is mainly driven by the following three real-world forces:
[0047] Upgraded Decision Analysis Needs: Enterprise decision-makers are no longer satisfied with basic data querying and report generation, but require the system to proactively identify anomalies from the data, predict trends, and provide decision-making basis;
[0048] The complexity of enterprise data environments is increasing dramatically: modern enterprise data structures exhibit three characteristics: heterogeneity, fragmentation, and real-time nature, requiring increasingly complex multidimensional analysis to reduce costs;
[0049] Traditional analytics processes are showing significant bottlenecks: the traditional workflow from data preparation to decision output suffers from multiple efficiency losses. Industry research indicates that data engineers spend 60% of their time cleaning and transforming data, while analysts dedicate 40% of their time to basic data verification rather than in-depth insight mining. This labor-intensive work model has become a major obstacle to enterprises' intelligent transformation.
[0050] Against this backdrop, intelligent data analysis devices based on Multi-Agent have emerged, with core application scenarios including:
[0051] Intelligent enterprise decision-making: Building enterprise-level "data brains" in fields such as the Internet, energy, and retail to support complex decisions ranging from supply chain optimization to sales forecasting.
[0052] Professional domain knowledge services: In scenarios such as updating clinical guidelines and interpreting legal provisions in the medical field, it is necessary to coordinate domain expert agents, evidence retrieval agents, and risk assessment agents to jointly ensure the professionalism and reliability of the output.
[0053] Based on this, embodiments of the present invention provide a data analysis method. (Refer to...) Figure 1 This invention provides a data analysis method, including:
[0054] S11, Obtain the advertising data to be analyzed and the questions expressed in natural language. The questions are used to express the analysis needs for the advertising data to be analyzed.
[0055] S12, determine the analysis type corresponding to the analysis requirement expressed by the question; based on the functional description information of multiple agents, select multiple agents that match the analysis type and the calling order of each agent as the target configuration information;
[0056] S13: According to the target configuration information, multiple agents are called in sequence to process the advertising data to be analyzed and obtain the data analysis results.
[0057] In this embodiment of the invention, by acquiring advertising data to be analyzed and a question expressed in natural language, the question represents the analysis requirement for the advertising data to be analyzed. The analysis type corresponding to the analysis requirement represented by the question is determined. Based on the functional description information of multiple agents, multiple agents matching the analysis type and the calling order of each agent are selected as target configuration information. This determines the target configuration information corresponding to the user's question expressed in natural language. The target configuration information includes the combination and calling order of multiple intelligent agents. Thus, multiple agents can be called sequentially according to the target configuration information to process the advertising data to be analyzed and obtain data analysis results. Each agent performs its own function. The combined action of multiple agents can reduce the exponentially accumulating error rate caused by a single LLM simultaneously handling data retrieval, logical reasoning, and conclusion generation, thereby improving the accuracy of advertising data analysis. This improvement in accuracy is particularly significant when facing complex tasks.
[0058] Furthermore, a single agent is prone to "professional blind spots" when processing cross-domain knowledge, resulting in contextual comprehension bias. Multiple agents, each with their own corresponding capabilities, reduce the possibility of agents handling cross-domain knowledge. Thus, the combined action of multiple agents can reduce contextual comprehension bias and further improve the accuracy of advertising data analysis.
[0059] Furthermore, the inherent "illusion" problem of LLM is amplified in data analysis scenarios. When training data includes test samples, the model directly outputs the remembered results instead of the actual reasoning conclusions, resulting in a "fabricated lie" phenomenon. Embodiment 1 of this invention constructs a data analysis framework based on the multiagent approach to overcome the data illusion caused by a single agent.
[0060] A single agent struggles to dynamically balance response speed and analytical depth, resulting in poor task adaptability. This invention improves task adaptability by combining multiple agents. Furthermore, when adding new data sources or analytical tools, a single agent requires global retraining, while enterprise-level systems often integrate hundreds of specialized tools, leading to difficulties in tool expansion. This invention addresses this by having multiple agents run independently. When new data sources or analytical tools are added, the relevant agents can be specifically expanded and trained, reducing the difficulty of expansion.
[0061] The data analysis method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a server, etc.
[0062] The data analysis method provided in this embodiment of the invention can be understood as an intelligent data analysis method based on multi-agent.
[0063] In S11, advertising data can include various data related to advertising. The advertising data to be analyzed can be represented in any form, such as document data or data in table format. In one example, the advertising data to be analyzed can include multiple data tables, each showing the daily revenue of the advertisement, as shown in Table 1. It should be noted that many factors influence advertising revenue, and the specific factors can be determined based on the actual situation, such as ad placement, platform, business line, project team, etc. The content shown in Table 1 is only an example.
[0064] Table 1
[0065] Ad space 1 Ad space 2 …… Ad space n Platform 1 Income 1 Income 4 …… Income 7 Platform 2 Income 2 Income 5 …… Income 8 …… …… …… …… …… Platform m Income 3 Income 6 …… Income 9
[0066] The question is used to express the analytical needs regarding the advertising data to be analyzed. For example, a question could be: "Why did the revenue from this advertisement decrease?" The question can be entered by the user based on their actual needs.
[0067] In S12, multiple intelligent agents can be pre-configured, and corresponding roles and tasks can be assigned to different agents. Simply put, different agents have their corresponding computing capabilities. For each agent, information is input to that agent, and the agent is invoked, so that the agent can obtain the processing results through its capabilities.
[0068] In this embodiment of the invention, multiple pre-configured agents can be assembled and connected in series to meet the analysis requirements expressed by the question. For example, the combination and calling order of intelligent agents corresponding to the analysis requirements can be selected from the pre-configured agents to complete the calculations required for the analysis requirements, realize data analysis for specific analysis requirements, and obtain the corresponding data analysis results.
[0069] The target configuration information may include all agents from a pre-configured pool of agents, or only a subset of them. Furthermore, the same agent may be included multiple times in the target configuration information.
[0070] In one optional embodiment, if the analysis type corresponding to the question indicates advertising revenue analysis, then the target pre-configuration information includes the combination and invocation order of multiple intelligent agents corresponding to the analysis strategy used for advertising revenue analysis.
[0071] Multiple pre-configured agent function descriptions can be provided. Thus, after determining the analysis type corresponding to the analysis requirement represented by the question, the agent that matches the required analysis type and the order in which each agent is called can be selected based on the pre-configured agent function descriptions.
[0072] In this process, semantic recognition can be performed on the question, and the result of the semantic recognition can indicate the type of analysis required for the question, that is, the analysis type corresponding to the question.
[0073] In one example, key fields can be extracted from the question to indicate the analysis type corresponding to the question. By matching these key fields with the functional descriptions of multiple pre-configured agents, it is possible to determine which agents match the analysis requirements with the desired analysis type, as well as the order in which these agents are invoked.
[0074] In one possible implementation, such as Figure 2 As shown, S12 includes:
[0075] S21, using a question classifier, determine the type of target analysis strategy corresponding to the question;
[0076] S22, find the correspondence between the configuration information corresponding to different analysis strategy types, and obtain the target configuration information corresponding to the target analysis strategy type.
[0077] The question classifier is used to categorize questions to determine the analysis type corresponding to the expressed analytical needs, that is, to determine the required analysis type for the question. The target analysis strategy type is the analysis type corresponding to the analytical needs expressed by the question.
[0078] Based on the functional descriptions of multiple agents, multiple agents matching the analysis type and the order in which each agent is invoked are selected as the target configuration information. This may include: finding the correspondence between the configuration information of different analysis strategy types to obtain the target configuration information corresponding to the target analysis strategy type.
[0079] Problem classifiers can be implemented using LLM (Local Management Model), and can also be called LLM agents.
[0080] In one possible approach, the mapping relationship between configuration information corresponding to different analysis strategy types can be pre-configured based on the functional description information of multiple agents. Thus, after determining the target analysis strategy type corresponding to the question through a question classifier, the mapping relationship between configuration information for different analysis strategy types can be found to obtain the target configuration information corresponding to the target analysis strategy type.
[0081] The correspondence between configuration information for different analysis strategy types can be established by mapping the identification information of different analysis strategy types to their corresponding configuration information.
[0082] The identification information for the analysis strategy type can be the name of the analysis strategy or a brief description of the analysis strategy, which indicates the function that the analysis strategy is to achieve.
[0083] The configuration information corresponding to the analysis strategy type can include the combination and calling order of multiple intelligent agents that implement the analysis strategy. Specifically, it can be represented by a flowchart, which shows which agents are called in sequence.
[0084] By finding the correspondence between the configuration information corresponding to different analysis strategy types, the target configuration information corresponding to the target analysis strategy type can be obtained. Before S12, the configuration information corresponding to each analysis strategy type can be pre-set and the correspondence between the configuration information corresponding to different analysis strategy types can be saved. In this way, after determining the target analysis strategy type corresponding to the question, the correspondence between the configuration information corresponding to different analysis strategy types can be found to obtain the target configuration information corresponding to the target analysis strategy type.
[0085] In one possible approach, pre-setting the configuration information corresponding to each analysis strategy type and saving the correspondence between the configuration information of different analysis strategy types can be achieved in the following way: Figure 3As shown, it includes:
[0086] S31, for each analysis strategy type, determine the identifiers of multiple pre-configured agents;
[0087] S32, receive the selection of identifiers for multiple pre-configured agents and determine the selection order of agents, determine the calling order of multiple agents corresponding to the analysis strategy type, and save the combination and calling order of multiple agents corresponding to the analysis strategy type.
[0088] For example, a user interface can be provided that displays identification information for multiple analysis strategy types for users to choose from. These users can be professional data analysts, etc.
[0089] When a trigger, such as a click, is detected that identifies an analysis strategy type, it indicates that the configuration information corresponding to that analysis strategy is being set. At this point, the identifiers of multiple pre-configured agents can be displayed on the user interface; these identifiers may include agent function descriptions, etc.
[0090] Furthermore, when the selection of identifiers for multiple pre-configured agents is received and the selection order of agents is determined, the invocation order of multiple agents corresponding to the analysis strategy type can be determined, and the combination and invocation order of multiple agents corresponding to the analysis strategy type can be saved.
[0091] In this embodiment, the analysis strategy type can be understood as a general term for a large model, which is composed of multiple agents. The multiple agents that make up the large model have a calling order, that is, which agent is called first and which agent is called later.
[0092] In another possible implementation, in this embodiment of the invention, in addition to inputting the question, a prompt can also be input. This prompt specifies the required analysis; it can be understood that the prompt can be used to parse the target configuration information corresponding to the question. For example:
[0093] The prompt is as follows:
[0094] You are a senior data analyst specializing in dimensional attribution analysis and data insights. A user will provide a tree-structured data report generated through dimensional attribution drill-down analysis. You need to assist the user in:
[0095] 1. **Data Validation**:
[0096] - First, check if the input content is a valid tree structure report.
[0097] - If invalid, output only: "Please pass in the correct data report".
[0098] 2. **Data Analysis**:
[0099] - Iterate through and analyze the changes in metrics at each node.
[0100] - Identify the key dimensions that influence the rise or fall of overall node metrics.
[0101] 3. **Report Optimization**:
[0102] - Maintain the original tree structure.
[0103] - Add a marker at the end of the key dimension nodes, in the format: [Major Rising Contributors] or [Major Falling Contributors]
[0104] - Tagging rules:
[0105] *Select 1-2 sub-dimensions that have the greatest impact on changes in the indicator.
[0106] 4. **Output Requirements**:
[0107] - Output format is HTML
[0108] - First, present the optimized, complete tree-structured report, without a header.
[0109] - Finally, add an "Analysis and Summary" section, which includes:
[0110] *Overall indicator trends
[0111] *Analysis of Key Influencing Factors
[0112] * Notable anomalies
[0113] - Strictly maintain the original data unchanged
[0114] 5. **Input Data Format Instructions:**
[0115] - Root node: Overall indicator status
[0116] -Child nodes: Metric information for a specified dimension in the drill-down analysis.
[0117] - Node format: Dimension description, Metric description
[0118] -Indicator description format: <Indicator 1>, <Indicator 2>,... (The first indicator is the primary analysis indicator)
[0119] Please analyze the following data report:
[0120] {{content}}
[0121] Here, "content" refers to the advertising data to be analyzed.
[0122] In S13, multiple agents can provide call interfaces. Once the target configuration information is determined, multiple agents can be called sequentially according to the target configuration information to process the advertising data to be analyzed and obtain the data analysis results.
[0123] The data analysis results can be presented in the form of an analysis report, such as a document.
[0124] In one optional embodiment, after obtaining the data analysis results in S13, the method may further include: returning the data analysis results to the user terminal.
[0125] Specifically, the server can send the data analysis results to the user's device. Furthermore, it can provide preview and download functions for the user to choose from.
[0126] The user terminal can be the terminal device used by the user. Specifically, the terminal device can have a client installed on it.
[0127] Referring to the above description, in this embodiment of the invention, multiple agents are pre-configured, and corresponding roles and tasks are assigned to different agents. In this way, multiple pre-configured agents can be assembled and connected according to analysis needs, so as to complete the calculations required for analysis needs, realize data analysis for specific analysis needs, and obtain corresponding data analysis results.
[0128] Multiple agents can be selected based on actual needs and other factors.
[0129] In one alternative embodiment, the present invention provides an advertising data analysis platform, which is deployed on a server and pre-configured with multiple agents for advertising data analysis.
[0130] In one possible implementation, the multiple agents may include at least two of the following: metric group attribution agent, dimension group attribution agent, dimension drill-down agent, anomaly analysis agent, report generation agent, and customer business logic attribution agent;
[0131] Among them, the attribution agent is used to identify core contribution metrics from the metrics that affect advertising revenue. The metrics that affect advertising revenue include: daily active users (DAU), average inventory per user, inventory fill rate, and advertising revenue per thousand impressions (ecpm).
[0132] Dimension group attribution agents are used to identify the core influencing dimensions from those that affect advertising revenue;
[0133] Dimension drill-down agents are used to determine the fine-grained impact of dimensions on data by decomposing the dimension hierarchy layer by layer;
[0134] Anomaly analysis agents are used to identify outliers in advertising data.
[0135] The report production agent is used to analyze, summarize, and polish the reports.
[0136] The customer business logic attribution agent is an attribution agent customized based on user needs.
[0137] The general logic of data analysis is broken down into different stages, such as indicator group attribution, dimension group attribution, dimension drill-down analysis, anomaly analysis, and customer business logic attribution. For each stage, an independent agent tool is built to produce various fine-grained analysis templates required for analysis reports.
[0138] Taking ad development as an example, ad revenue = DAU * average inventory per user * inventory fill rate * ecpm. The specific calculation rules for each agent are as follows:
[0139] Attribution agent for indicator groups:
[0140] Revenue fluctuations directly impact four key metrics: DAU, average inventory per employee, inventory fill rate, and eCPM. Which metric is the core contributor to revenue? This is addressed using an attribution contribution algorithm. The multiplicative formula can be transformed into an additive formula using natural exponentiation.
[0141] The contribution value of each indicator is determined based on the indicator attribution contribution algorithm; indicators whose contribution values meet the preset conditions are selected as core contribution indicators.
[0142] Indicator contribution value = Difference in core indicators * log(multiplication factor of comparison date / multiplication factor of base date) / log(core indicator of comparison date / core indicator of base date)
[0143] The core contribution index can be obtained by calculating the contribution value of the indicators, thereby inferring which factor plays a key role in the fluctuation of market income.
[0144] Dimension group attribution agent and dimension drill-down agent:
[0145] The dimensional attribution algorithm here uses the explanatory power (EP value) from the multidimensional time series anomaly root cause analysis algorithm (Adtributor) to represent the contribution. The agent is responsible for taking the core indicators and the dimensions to be attributed as input, and outputting the normalized contribution of the dimension factors. Dimension group attribution is represented by the cumulative contribution of multiple dimension attribution contributions (variance explanation). Multiple dimension contribution rankings are obtained at once, and drill-down is performed based on the top dimension to obtain the main contributing factors.
[0146] Anomaly analysis agent:
[0147] This invention employs the Prophet algorithm for time series data anomaly detection. This algorithm can handle outliers and missing values in time series data and can automatically detect the effects of periodic trends and holidays. The Prophet algorithm is based on a decomposable (trend + season + holiday) model, and its input includes a ds column (time) and a y column (numerical variable, i.e., the predicted value). It is suitable for time series with strong periodicity and relatively regular fluctuations.
[0148] Of course, the system time can be extended to include more anomaly detection algorithms, such as Holt-winters.
[0149] Anomaly analysis is mainly used to identify outliers in time series models. Agents are mainly used to actively explore and discover outliers to trigger more advanced attribution calculations.
[0150] Traditional architectures lack cross-validation, often directly transmitting erroneous results to users, indicating a missing verification mechanism. Furthermore, anomaly detection models lacking cross-validation may misinterpret data collection errors as market fluctuations, leading to flawed decision-making. This invention, through an anomaly analysis agent, compensates for this lack of a verification mechanism by implementing anomaly detection, further improving the accuracy of advertising data analysis and reducing erroneous decisions arising from data analysis results.
[0151] Customer business logic attribution agent:
[0152] The business logic here can be customized according to user needs. The system has developed an interface mode that supports two models: user-defined Structured Query Language (SQL) and Python (a programming language) scripts. It can expand data acquisition capabilities to obtain more customized and easy-to-use capabilities.
[0153] Each agent is responsible for a specific computing capability, and they can be assembled and linked together to form a comprehensive analysis of a particular business scenario. For example, the combination pattern is: anomaly analysis agent --> metric group attribution agent --> dimension drill-down agent 1 --> dimension drill-down agent 2 --> dimension drill-down agent n.
[0154] The final analysis report is derived by combining the results returned by each agent. To enhance report parsing and improve writing quality, this embodiment of the invention provides a "report generation agent" for summarizing and refining the analysis report.
[0155] This invention provides a multi-agent data analysis framework that leverages the inductive reasoning capabilities of LLM (Limited Modeling) to assign roles and tasks to different agents and equip them with corresponding tool plugins, thereby completing complex tasks. Key capabilities include dimension group attribution agents, dimension drill-down agents, indicator group attribution agents, anomaly analysis agents, and customer business analysis agents. Multiple agent capabilities can be combined according to actual needs to complete data attribution and core node data collection. Finally, the framework utilizes large-scale model text template organization capabilities to optimize the output of analysis reports.
[0156] In one example, such as Figure 4 As shown, the outermost agent adopts the general LLM agent model, namely the LLM + knowledge base model, which allows users to ask questions in natural language. Through the question classifier, different downstream agents are selected to collect data, perform attribution analysis, and aggregate results. Finally, the report production agent summarizes and refines the report and outputs it to the user.
[0157] For example, first provide an input window, and then display "Start" in the input window to provide input of advertising data to be analyzed and questions to be asked;
[0158] Upon receiving the advertising data to be analyzed and the questions asked, a classifier can be used to determine which analysis model to employ. For example, analysis model 1 is used for category 1, analysis model 2 for category 2, analysis model 3 for category 3, analysis model 4 for category 4, analysis model 5 for category 5, and other analysis models for category 6. This embodiment of the invention does not limit the specific form of the classifier; for example, it can be the language model (GPT-4o) released by OpenAI for the chatbot ChatGPT.
[0159] Different analysis models can be combined with corresponding knowledge bases. For example, analysis model 1 can be used for knowledge retrieval, and the knowledge base of advertising data TEST and proprietary exposure data.txt can be combined during the knowledge retrieval process; analysis model 2 can be used for traffic analysis, and the data text ad_data.txt can be combined during the traffic analysis process; analysis model 3 can be used for revenue analysis, and the data text ad_data.txt can be combined during the revenue analysis process, and so on.
[0160] When implementing knowledge retrieval, specific analyses can be performed on traffic, budget, revenue, and anomaly detection, for example, using the GPT-4o model. After these analyses, a response can be directly generated. When implementing traffic analysis, specific analyses can be performed on year-on-year and month-on-month revenue growth and inventory occupancy, which can also be implemented using the GPT-4o model. During revenue analysis, specific analyses can be performed on customer cost deviations, also using the GPT-4o model. After year-on-year and month-on-month revenue growth, inventory occupancy, or customer cost deviation analyses, a report production agent can be invoked to generate data analysis results, which can be understood as a data analysis report. The report production agent can also be implemented using the GPT-4o model.
[0161] Each analysis agent is backed by multiple attribution analysis modules that can be freely combined and configured. Taking revenue year-on-year and month-on-month fluctuation analysis as an example, such as... Figure 5 As shown:
[0162] When triggering the start of year-on-year and month-on-month revenue fluctuation analysis, you can first perform dimensional attribution - business line, indicator group attribution: in-site revenue + affiliate, dimensional attribution - [in-site industry], dimensional attribution - [affiliate SKP], and dimensional attribution - [customer].
[0163] Figure 5 It can also be understood as a representation of the configuration information corresponding to the above analysis strategy types.
[0164] The data analysis method provided in this invention abandons the popular ChartBI SQL generation model and starts with specific attribution analysis algorithms to reduce the data collection illusion and error rate of large models, ensuring the accuracy and ease of use of the analysis report. It adopts a decision tree + contribution attribution algorithm to simplify the data collection process, and multiple verification mechanisms avoid the risks caused by data errors. The verification mechanisms refer to various attribution calculation capabilities during the data acquisition process; this part is implemented via an interface and can be customized, avoiding the model illusion problem by not using large model production.
[0165] Specifically, the decision tree can include combinations and invocation order of multiple intelligent agents represented by the above configuration information. For example, as mentioned above... Figure 5 The flowchart shown illustrates the decision-making logic.
[0166] Multi-Agent-based intelligent data analysis devices overcome technical bottlenecks through a distributed cognitive architecture. By specializing agents, they solve cognitive overload and improve the processing speed of complex tasks by 300%. Minute-level response replaces day-long manual processes, reconstructing decision-making efficiency. This achieves a revolutionary leap in the analysis paradigm from "manual verification" to "autonomous and trustworthy".
[0167] Based on the same inventive concept as the data analysis method described above, embodiments of the present invention also provide a data analysis device, such as... Figure 6 As shown, it includes:
[0168] The acquisition module 601 is used to acquire the advertising data to be analyzed and the questions expressed in natural language. The questions are used to express the analysis needs for the advertising data to be analyzed.
[0169] The determination module 602 is used to determine the analysis type corresponding to the analysis requirement represented by the question; based on the functional description information of multiple agents, it selects multiple agents that match the analysis type and the calling order of each agent as the target configuration information;
[0170] The analysis module 603 is used to sequentially call multiple agents according to the target configuration information to process the advertising data to be analyzed and obtain the data analysis results.
[0171] Optionally, the determining module 602 is specifically used to determine the target analysis strategy type corresponding to the question through the question classifier; find the correspondence between the configuration information corresponding to different analysis strategy types, and obtain the target configuration information corresponding to the target analysis strategy type.
[0172] Optionally, the device further includes:
[0173] The configuration module is used to determine the identifiers of multiple pre-configured agents for each analysis strategy type before searching for the correspondence between configuration information for different analysis strategy types; receive the selection of the identifiers of the multiple pre-configured agents and determine the selection order of the agents, determine the calling order of the multiple agents corresponding to the analysis strategy type, and save the combination and calling order of the multiple agents corresponding to the analysis strategy type.
[0174] Optionally, the multiple agents include at least two of the following: indicator group attribution agent, dimension group attribution agent, dimension drill-down agent, anomaly analysis agent, and customer business logic attribution agent;
[0175] Among them, the attribution agent is used to identify core contribution metrics from the metrics that affect advertising revenue. The metrics that affect advertising revenue include: daily active users (DAU), average inventory per user, inventory fill rate, and advertising revenue (ecpm) per thousand impressions.
[0176] Dimension group attribution agents are used to identify the core influencing dimensions from those that affect advertising revenue;
[0177] Dimension drill-down agents are used to determine the fine-grained impact of dimensions on data by decomposing the dimension hierarchy layer by layer;
[0178] Anomaly analysis agents are used to identify outliers in advertising data;
[0179] The customer business logic attribution agent is an attribution agent customized based on user needs.
[0180] Optionally, the target pre-configuration information includes the combination and invocation order of multiple intelligent agents corresponding to the analysis strategy used for advertising revenue analysis.
[0181] This invention also provides an electronic device, such as... Figure 7 As shown, it includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other through the communication bus 704.
[0182] Memory 703 is used to store computer programs;
[0183] The processor 701 is used to execute the program stored in the memory 703 to implement the above-described data analysis method steps.
[0184] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.
[0185] The communication interface is used for communication between the aforementioned terminal and other devices.
[0186] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0187] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0188] In another embodiment of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements any of the data analysis methods described in the above embodiments.
[0189] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the data analysis methods described in the above embodiments.
[0190] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).
[0191] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0192] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, electronic devices, computer-readable storage media, and computer program products are basically similar to the method embodiments, and therefore the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0193] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. A data analysis method, characterized in that, include: Acquire advertising data to be analyzed and a question expressed in natural language, wherein the question is used to express the analysis needs for the advertising data to be analyzed; Determine the analysis type corresponding to the analysis requirement represented by the question; Based on the functional description information of multiple agents, select multiple agents that match the analysis type and the calling order of each agent as the target configuration information; According to the target configuration information, multiple agents are invoked sequentially to process the advertising data to be analyzed, and the data analysis results are obtained.
2. The method according to claim 1, characterized in that, The process involves determining the analysis type corresponding to the analysis requirement represented by the question; selecting the agent that matches the required analysis type and the calling order of each agent based on the pre-configured functional description information of multiple agents, as the target configuration information, including: The target analysis strategy type corresponding to the question is determined by the question classifier; Find the correspondence between the configuration information corresponding to different analysis strategy types to obtain the target configuration information corresponding to the target analysis strategy type.
3. The method according to claim 2, characterized in that, Before finding the correspondence between configuration information for different analysis strategy types, the method further includes: For each analysis strategy type, identify the identifiers of multiple pre-configured agents; The system receives the selection of identifiers for multiple pre-configured agents and determines the selection order of the agents, determines the invocation order of multiple agents corresponding to the analysis strategy type, and saves the combination and invocation order of multiple agents corresponding to the analysis strategy type.
4. The method according to claim 1, characterized in that, Multiple agents include at least two of the following: metric group attribution agent, dimension group attribution agent, dimension drill-down agent, anomaly analysis agent, and customer business logic attribution agent; The attribution agent is used to determine the core contribution metrics from the metrics that affect advertising revenue. The metrics that affect advertising revenue include: daily active users (DAU), average inventory per user, inventory fill rate, and advertising revenue (ecpm) per thousand impressions. The dimension group attribution agent is used to identify the core influencing dimensions from those that affect advertising revenue; The dimension drill-down agent is used to determine the fine-grained impact data of dimensions through the layer-by-layer decomposition of the dimension hierarchy structure; The anomaly analysis agent is used to identify anomalies in the advertising data; The customer business logic attribution agent is an attribution agent customized according to user needs.
5. The method according to any one of claims 1 to 4, characterized in that, The target pre-configuration information includes the combination and invocation order of multiple intelligent agents corresponding to the analysis strategy used for advertising revenue analysis.
6. A data analysis device, characterized in that, include: The acquisition module is used to acquire advertising data to be analyzed and questions expressed in natural language, wherein the questions are used to express the analysis needs for the advertising data to be analyzed; The determination module is used to determine the analysis type corresponding to the analysis requirement represented by the question; and select multiple agents that match the analysis type and the calling order of each agent as target configuration information based on the functional description information of multiple agents. The analysis module is used to sequentially call multiple agents according to the target configuration information to process the advertising data to be analyzed and obtain data analysis results.
7. The apparatus according to claim 6, characterized in that, The determining module is specifically used to determine the target analysis strategy type corresponding to the question through a question classifier; find the correspondence between the configuration information corresponding to different analysis strategy types, and obtain the target configuration information corresponding to the target analysis strategy type.
8. The apparatus according to claim 7, characterized in that, The device further includes: The configuration module is used to determine the identifiers of multiple pre-configured agents for each analysis strategy type before searching for the correspondence of configuration information corresponding to different analysis strategy types; receive the selection of the identifiers of the multiple pre-configured agents and determine the selection order of the agents, determine the calling order of the multiple agents corresponding to the analysis strategy type, and save the combination and calling order of the multiple agents corresponding to the analysis strategy type.
9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-5.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method described in any one of claims 1-5.