A four-stage automatic attribution method and system for enterprise index system

Abstract

The application discloses a four-stage automatic attribution method and system for an enterprise index system, which standardizes attribution into four progressive analyses. In the first stage, Z-score and persistence double-checking are used to detect index abnormalities and distinguish true abnormalities from random fluctuations. In the second stage, multi-dimensional drilling decomposition is carried out, and dimension contribution ranking is output according to dimension explanatory power. In the third stage, structural, efficiency and cross-effect three-component variance decomposition is carried out on high-explanatory-power dimensions, and the contributions of the two types of factors to index abnormality are quantified. In the fourth stage, an index cause-effect graph is relied on to output a root cause list with confidence by weighted reverse source tracing and traversal. The application realizes dynamic iterative optimization of the cause-effect graph through attribution feedback, generates a structured attribution report supporting interactive drilling, and solves the problems of low attribution efficiency of enterprise indexes, mixed superposition of structural efficiency factors, and difficult accumulation and optimization of attribution capability.

Description

Technical Field

[0001] This invention belongs to the field of data analysis and business intelligence technology, specifically relating to a four-stage automatic attribution method and system for enterprise indicator systems. It is particularly suitable for intelligent decision analysis platforms that need to perform automated root cause analysis on abnormal changes in business indicators and provide quantifiable attribution explanations. Background Technology

[0002] In the daily operation and management of enterprises, when core business indicators (such as sales, order volume, user retention rate, conversion rate, etc.) experience abnormal fluctuations, managers and data analysts need to quickly and accurately pinpoint the root causes of the changes in order to take timely countermeasures. However, in complex indicator systems with hundreds or even thousands of interrelated metrics, manual attribution presents the following serious challenges: (1) Dimensional space explosion: Enterprise business indicators are usually associated with dozens of analysis dimensions (region, channel, category, store, salesperson, etc.), and the number of members in each dimension ranges from a few to thousands. It is completely infeasible in terms of time to exhaustively analyze all the combination of dimensions. (2) Complex relationships between indicators: There are multiple complex relationships between business indicators, such as computational dependence (composite indicators referencing sub-indicators) and business impact (upstream operational indicators affect downstream result indicators), making it difficult to systematically trace the source by relying solely on human experience; (3) Overlapping structural and efficiency changes: Changes in business indicators often include both structural factors (such as a decrease in the proportion of high-profit users) and efficiency factors (such as a decrease in per capita consumption of each segment group). Manual analysis makes it difficult to quantitatively distinguish the contribution of the two types of factors. (4) High subjectivity and low repeatability of attribution: Different analysts have significantly different attribution conclusions for the same abnormal indicator, and the attribution process lacks standardized records, making it impossible to systematically reuse historical attribution experience. The main solutions in the existing technology include: Option 1: Manual drill-down analysis. Analysts manually drill down by dimension using BI tools to examine the contribution of each dimension one by one. This option relies on personal experience, is inefficient, cannot cover all dimension combinations, and has a high rate of attribution omissions. Option 2: Rule-based alarm system. This system uses pre-set fixed rules (e.g., triggering an alarm when a member's fluctuation exceeds a threshold in a certain dimension). However, this approach has high rule maintenance costs, only covers known anomaly patterns, is ineffective against novel anomalies, and cannot provide attribution explanations. Option 3: General machine learning attribution methods (such as SHAP values, feature importance). This approach models the metric prediction and approximates the attribution contribution with feature importance. However, this option requires a large amount of labeled training data, has insufficient model interpretability, and does not consider the semantic structure of the business metric system (metric causal relationship graph), making the attribution results unintuitive for business personnel. None of the above solutions combined the systematic nature of multi-dimensional drill-down decomposition with the quantitative accuracy of variance decomposition, nor did they utilize the causal graph structure of the indicator system itself for root cause tracing, resulting in attribution efficiency and quality that are difficult to meet the actual needs of enterprises. Summary of the Invention

[0003] Purpose of the invention The purpose of this invention is to provide a four-stage automatic attribution method and system for enterprise indicator systems. It organically integrates anomaly detection, multi-dimensional drill-down decomposition, variance decomposition contribution quantification, and root cause tracing of indicator causal graphs into a standardized four-stage process. This enables fully automatic, quantifiable, and interpretable attribution output for complex business indicator anomalies, solving technical problems such as low attribution efficiency, overlapping structural and efficiency factors, and inability to reuse historical experience in existing technologies. Technical solution Core Approach: The attribution problem is broken down into four progressively deeper analytical stages. The output of each stage provides targeted input for the next, ultimately culminating in a list of explainable root causes that includes quantified contribution values ​​and confidence scores. Simultaneously, through incremental updates to the causal graph of indicators and an attribution feedback reinforcement mechanism, attribution capabilities are continuously improved as business experience accumulates. Mechanism 1: Anomaly Detection Based on Statistical Significance Screening The system employs a dual verification mechanism of Z-score and persistence to accurately distinguish between statistically significant anomalies and random fluctuations, avoiding resource waste caused by erroneous attribution tasks, while ensuring that genuine business anomalies are not overlooked. Mechanism 2: Multi-dimensional drill-down based on dimensional explanatory power ranking The contribution of all related dimensions is systematically calculated, and the concept of "dimensional explanatory power" is introduced to quantify the ability of each dimension to explain the changes in the indicators. This provides priority analysis dimensions for subsequent variance decomposition and avoids incomplete attribution due to omission of analysis dimensions. Mechanism 3: Variance decomposition of structural and efficiency effects By introducing the structure-efficiency decomposition framework from economics, the changes in indicators are precisely broken down into three components: structural effect, efficiency effect, and cross-effect, thus quantitatively answering the core business question: "Is the change caused by changes in business structure or efficiency?" Mechanism 4: Tracing the Root Causes of the Integration of Knowledge Graphs and Statistical Associations The system integrates the dependence of indicator calculation (hard association), business logic association (expert knowledge), and historical statistical correlation (data mining) into a unified indicator causal graph. It performs a reverse traversal of causal association strength weighted on the graph and outputs the root cause ranking with confidence scores, so that the attribution conclusions are reasonable and verifiable. Beneficial effects (1) Significantly improve attribution efficiency: The fully automated four-stage process compresses the manual attribution that originally required analysts to several hours to minutes, and covers all related dimensions without any risk of omission. (2) Quantitative attribution contribution: Each attribution conclusion is accompanied by a quantitative contribution value and percentage, which allows managers to intuitively assess the impact of each root cause and make priority decisions. (3) Precise distinction between structural and efficiency factors: The variance decomposition mechanism clearly distinguishes between structural effects and efficiency effects, providing a basis for decision-making for two completely different coping strategies: "optimizing structure" and "improving efficiency". (4) Self-evolution of attribution ability: The indicator causal graph strengthens continuous learning through attribution feedback, and historical attribution experience is transformed into graph weight accumulation, so that the system’s attribution accuracy for similar anomalies continues to improve over time. Attached Figure Description Figure 1 This is a flowchart of the four-stage automatic attribution method, showing the logical relationship and main inputs and outputs of the four stages: anomaly detection, multi-dimensional drill-down decomposition, variance decomposition contribution calculation, and root cause tracing of index graphs. Figure 2 This is a schematic diagram of multi-dimensional drill-down decomposition, demonstrating an example of the calculation process for the absolute contribution of each member in each dimension and the ranking of dimension explanatory power. Figure 3 This is a schematic diagram of the three components of variance decomposition, which shows the quantitative results of the total change in the index being decomposed into three types of components: structural effect, efficiency effect, and cross effect. Figure 4 This is a schematic diagram of the causal graph structure, showing a multi-layered causal graph composed of three types of edges (computational dependency edges, business logic edges, and statistical correlation edges) and a reverse root cause tracing path. Figure 5 This is a flowchart of attribution feedback reinforcement and graph iteration update, showing the positive reinforcement, negative adjustment mechanism, and decay mechanism of graph weights after the user confirms the root cause. Detailed Implementation Example 1: Attribution Scenario for Abnormal Decline in "Daily Sales" of Retail Enterprises On November 15, 2024 (Friday), a retail company's data platform detected that the "daily sales" indicator had fallen by 18.7% compared to the historical benchmark for the same period, with a Z score of -3.2 (exceeding the threshold of -2.5), and had been abnormal for two consecutive days, triggering an attribution task. Phase 1: Anomaly Detection Output Target metric: Daily sales Observed value: RMB 12.47 million (November 15, 2024) Historical benchmark (average of the same period in the last 30 days): 15.33 million yuan Deviation amount: -2.86 million yuan (-18.7%) Z-score: -3.2 (significantly abnormal) Duration: 2 days (trigger attribution) Phase Two: Multi-dimensional Drilling Decomposition The system iterates through the 8 dimensions associated with "daily sales," calculates the explanatory power of each dimension, and outputs the top 3 dimensions: Dimensional Explanatory Power Ranking: 1. Channel Dimension (Explanatory Power: 76.3%) - Offline store channels: -3.12 million yuan (primary contributor) - Online self-operated channels: contributed +260,000 yuan (positive) 2. Category Dimension (Explanatory Power: 68.1%) - Home Appliances Category: -1.98 million yuan - Apparel Category: Contribution -1.07 million yuan 3. Regional dimension (Explanatory power: 43.2%) East China Region: Contribution -1.65 million yuan - South China Region: Contribution -980,000 RMB Phase 3: Variance Decomposition Decomposition of the execution variance of -3.12 million yuan for offline store channels using the combined dimension of "channel × category": Structural effect: -890,000 yuan (31.1%) — The traffic share of high-priced product categories (home appliances) in offline channels decreased. Efficiency effect: -1.98 million yuan (63.5%) – The average transaction value per customer decreased across all product categories in offline channels. Cross-effect: -250,000 yuan (5.4%) — Cross-component of the combined effect of structure and efficiency. Dominant change type: Efficiency-driven (efficiency effect accounts for 63.5%) Conclusion: The decline in daily sales was mainly due to decreased efficiency (lower average order value), with structural changes (product category traffic structure) as a secondary factor. Phase 4: Root Cause Analysis of Indicator Maps Starting from the abnormal node in "daily sales," traverse backwards along the causal graph of the indicator to identify candidate root causes: Root cause ranking list: 1. Average transaction value at physical stores (Association strength: 0.87, confidence level: 91%) Impact path: Offline average order value ↓ → Offline daily sales ↓ → Daily sales ↓ Quantitative contribution: -1.98 million yuan (main reason) Recommendation: Examine the discount levels and changes in customer traffic quality during the same period of offline promotional activities. 2. Number of offline visitors (correlation strength: 0.61, confidence: 74%) Influence path: Decrease in number of offline visitors → Decrease in daily offline sales → Decrease in daily sales Quantified contribution: -780,000 yuan (secondary factor) Suggestion: Check weather in the same period, competitor activities, and abnormal store operations 3. Promotion intensity of home appliance category (correlation strength: 0.53, confidence: 68%) Influence path: Decrease in home appliance promotions → Decrease in average transaction value of home appliances → Decrease in daily sales Quantified contribution: -620,000 yuan (secondary factor) Example 2: Attribution feedback reinforcement and graph iteration scenario Based on the attribution result of Example 1, the operation analyst confirms that the real root cause of this sales decline is "the average transaction value falls back due to the fading residual effect of Double 11", marks and confirms the root cause as "average transaction value of offline stores" in the system, and denies that "number of offline visitors" is the root cause of this decline. Implementation of attribution feedback reinforcement: - The edge weight of "average transaction value of offline stores → daily sales" is positively reinforced by +0.05; - The edge weight of "number of offline visitors → daily sales" is slightly adjusted negatively by -0.01; - The timestamp of this reinforcement is recorded for subsequent attenuation calculation. After feedback reinforcement of 30 historical attribution tasks, the root cause positioning accuracy of the system for similar "efficiency-led offline channel sales decline" has increased from the initial 61% to 89%.

Claims

1. A four-stage automatic attribution method for enterprise indicator systems, characterized in that, The following four stages are executed sequentially: Phase 1: Anomaly Detection (1) Continuously monitor the index value sequence in a sliding time window manner, and calculate the historical benchmark value for each monitoring index. The historical benchmark value is the statistical mean of the data in the same period within a specified backtracking period (such as the past 30 natural days). (2) Calculate the deviation and deviation rate between the current observation and the historical benchmark. When the deviation rate exceeds the preset threshold range and the statistical significance meets the set conditions, mark the indicator as abnormal. The statistical significance is determined by the ratio of the deviation to the historical standard deviation (Z score). When the absolute value of the Z score exceeds the threshold, an anomaly marker is triggered. (3) Generate attribution tasks for indicators marked as anomalous, and proceed to the second stage; Phase Two: Multi-dimensional Drilling Decomposition (4) Obtain all dimension-type virtual fields associated with the target indicator, perform grouping and aggregation for each dimension, and calculate the indicator value and year-on-year change of each member value of each dimension in the current observation period and the comparison benchmark period; (5) For each dimension, calculate the "absolute contribution" (the sum of the changes of each member equals the total change of the indicator) and "contribution percentage" (the absolute value of the absolute contribution of each member divided by the absolute value of the sum of the absolute contributions of all members), sort them in descending order of contribution percentage, and output the Top-K contributing members list for each dimension. (6) For each dimension in step (5), calculate the "dimensional explanatory power" of that dimension to the change of the indicator: divide the sum of the absolute contributions of the Top-K members of that dimension by the absolute value of the total change of the indicator. The higher the explanatory power of the dimension, the stronger the ability of that dimension to explain the change. Output the dimensional sorting results in descending order of explanatory power, and proceed to the third stage; Phase 3: Calculation of Variance Decomposition Contribution (7) For the high explanatory power dimensions identified in the second stage, perform "multi-dimensional joint variance decomposition": construct a decomposition model with the change in the index as the dependent variable and the change in the structure and efficiency of the combination of members of each dimension as independent variables; (8) The total change in the indicator is decomposed into the following three components: Structure Effect: The contribution of changes in the proportion of members in each dimension to the index, that is, the change in the index caused solely by structural changes under the assumption that efficiency remains constant. Efficiency Effect: The contribution of unit efficiency changes of each dimension member to the index, that is, the index change caused solely by efficiency changes under the assumption that the structure remains unchanged. Interaction effect: The cross-component generated by the combined effects of structural and efficiency changes; (9) Output the quantitative contribution value of each component and its proportion of the total change, identify the dominant change type (structure-dominated / efficiency-dominated / hybrid), and provide directional attribution input for the fourth stage; Phase 4: Root Cause Analysis of Indicator Maps (10) Based on the "indicator causal graph" pre-constructed in the indicator system, the indicator causal graph takes indicators as nodes and business causal relationships (computational dependence, business logic association, historical attribution correlation) as directed edges, and starts from the abnormal indicator node and performs reverse tracing along the causal edges; (11) During the traversal, for each candidate upstream root cause node, calculate its “causal correlation strength” with the abnormal index: comprehensively consider the historical co-fluctuation correlation coefficient, the edge weight of the index graph and the dominant change type determined in step (8); (12) Select a set of candidate root cause nodes whose causal association strength exceeds the set threshold, and combine the dimensional explanatory power of step (6) and the variance decomposition results of step (8) to generate a "root cause ranking list". Each root cause includes: root cause index / factor identifier, influence path description, quantitative contribution value and confidence score. (13) Output the final attribution report, which includes: basic information of abnormal indicators, contribution analysis of top dimensions, variance decomposition component plot, root cause ranking list and suggested investigation priority.

2. The method according to claim 1, characterized in that, In the first stage step (2), the statistical significance judgment is further verified as follows: when the abnormal marker appears consecutively at N adjacent time points (N≥2) within the time window, it is considered a continuous abnormality. Single-point occasional deviations do not trigger the attribution task, so as to reduce false triggering caused by data fluctuations.

3. The method according to claim 1, characterized in that, In the second stage step (5), for dimensions where members are added or disappear (such as newly opened stores or closed stores), an additional "structural change correction amount" is calculated: for members that did not exist in the comparison period, their contribution is listed separately as "new member effect"; for members that did not exist in the current period, their contribution is listed separately as "disappeared member effect"; the two are not included in the contribution ranking of regular members, but are presented separately in the report as independent explanatory items.

4. The method according to claim 1, characterized in that, In the fourth stage step (11), the historical co-fluctuation correlation coefficient is calculated based on the sliding correlation window: in the past T statistical periods, the Pearson correlation coefficient of the change rate sequence of the candidate root cause index and the target index is calculated periodically, and its weighted average is taken (the recent period is given higher weight) to reflect the time-dependent decay characteristics of the causal relationship strength.

5. A method for constructing and maintaining a causal graph of enterprise indicators that supports incremental updates, used to support the root cause tracing in the fourth stage of claim 1, characterized in that, Includes the following steps: (1) Graph initialization: The initial structure of the index causal graph is constructed using the following three types of edge sources: Computational dependency edges: Directed edges are automatically extracted from the calculation template of the indicator system and established between composite indicators and their operand indicators, and between derived indicators and their basic atomic indicators. Business logic edge: Manually configured by business knowledge, describing the pairs of indicators that have a direct or indirect impact on the business (such as "customer acquisition cost" affecting "new customer GMV"). Statistical correlation edges: The system automatically mines from historical data and establishes weighted correlation edges for indicator pairs whose sliding correlation coefficients have been consistently higher than the threshold in historical data. (2) Incremental update: When the indicator system adds new indicators or adjusts the calculation logic, the calculation dependency edges are automatically re-extracted and the graph is updated; when enough historical data is accumulated (reaching the minimum number of statistical periods), the statistical correlation edge weights are recalculated and the graph is updated periodically. (3) Attribution feedback reinforcement: After each attribution task is completed and the root cause is confirmed by the user, the weight of each edge in the attribution path is positively reinforced (weight increases), and the candidate edges that do not appear in the final root cause path are slightly negatively adjusted, so that the graph weights are continuously optimized as actual attribution experience accumulates.

6. The method according to claim 5, characterized in that, The attribution feedback reinforcement in step (3) adopts a decay mechanism: historical reinforcement records are assigned decay weights according to time (the earlier the reinforcement record, the lower the weight), so as to avoid historical attribution experience from misleading the new attribution task after the business model changes, and to maintain the map's adaptability to business changes.

7. A method for generating structured attribution reports and interactive drill-down based on four-stage attribution results, characterized in that, Includes the following steps: (1) The system automatically generates a structured attribution report, which is organized in a hierarchical structure: the first layer is the indicator overview (abnormal summary, comparison of observed values ​​with benchmark values); the second layer is the dimensional contribution analysis (ranking of the explanatory power of each dimension, list of top members' contributions); the third layer is the variance decomposition diagram (visualization of the proportion of structural effects / efficiency effects / cross effects); the fourth layer is the root cause list (root cause ranking, description of the impact path, confidence level). (2) Each dimension member contribution item and root cause node in the report supports "interactive drill-down": When a user clicks on a specific dimension member, the system uses that member as a constraint to re-execute the second to fourth stages of attribution for the target indicator and generate a sub-attribution report from the perspective of that member. (3) The system supports users to mark, confirm or correct the attribution results: users can mark "confirmed root cause", "non-root cause" or add supplementary explanations; the marking results trigger the attribution feedback reinforcement process in step (3) of claim 5.

8. An automatic attribution system for enterprise metrics that implements the methods of claims 1 to 7, characterized in that, include: Anomaly detection module: continuously monitors indicator value sequences, calculates Z-scores, performs continuous verification, and generates attribution tasks; Multi-dimensional drill-down decomposition module: Traverses related dimensions, calculates the absolute contribution, contribution ratio, and explanatory power of each member, and outputs the dimension ranking; Variance decomposition module: Performs three-component variance decomposition of structural effects, efficiency effects, and cross effects, and outputs the quantitative contribution value of each component and the dominant change type; Indicator Causal Graph Module: Maintains the indicator causal graph (including three types of edges and weights), and supports incremental updates of the graph and attribution feedback reinforcement; Root Cause Tracing Module: Performs reverse traversal on the indicator causal graph, calculates the strength of causal relationships, and generates a root cause ranking list; Attribution report generation module: Integrates the output of the four stages to generate a hierarchical attribution report, supporting interactive drill-down and user annotation feedback.

9. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images