Comprehensive Attribution Analysis Methodology and Its Business Application

1. Introduction

Attribution analysis is a method that uses data and logical reasoning to identify which factors (product, price, market, etc.) cause specific outcomes such as business volume, conversion rate, or satisfaction, and how they influence those outcomes. It is applied in fields like psychology, investment, and online advertising to guide decision‑making and reduce blind experimentation.

2. Background

Business stakeholders frequently ask why conversion rates drop or why user churn increases. Effective attribution analysis closes the loop between business, data, and operations, enabling faster, higher‑quality, and more explainable insights.

3. Research

The traditional four‑step attribution process (data collection, statistical analysis, problem pinpointing) suffers from simple dimensions, low efficiency, difficulty in deep mining, and improper dimension handling, leading to high time consumption and inconsistent analysis quality.

3.1 Business Side

Improve data accuracy through data cleaning.

Boost basic analysis efficiency with standardized SQL and consolidated reports.

Enhance result interpretability by building a comprehensive business dimension system.

3.2 Algorithm Side

Increase analysis efficiency.

Improve analysis quality.

Establish a standard attribution workflow.

4. Solution

4.1 Overall Plan

A time‑series based fluctuation attribution workflow provides a complete, efficient analysis method that guides analysts from problem discovery to causal factor identification and recommendation.

4.2 Detailed Steps

2.1 Discover Issues from Business Data

Monitor alerts and reports to trigger abnormal fluctuation thresholds and define clear attribution goals.

2.2 Classify and Locate Problem Causes

Horizontal attribution uses multi‑branch trees and SQL to generate self‑service email reports; vertical attribution uses dashboard statistics to pinpoint conversion stages.

2.3 Qualify Problem Causes as Controllable

Identify controllable vs. uncontrollable factors (time and space). Time factors are handled with trend prediction; space factors are detected via Simpson's paradox.

2.4 Causal Detection for Controllable Factors

Apply Bayesian network‑based causal detection to reveal causal links and influence weights.

4.3 Highlights

3.1 Identifying Spatial Factors (Simpson's Paradox)

Simpson's paradox occurs when combined data yields opposite conclusions to separate groups; it helps detect user‑structure‑driven conversion fluctuations.

3.2 Automated Feature Selection

自动化和智能化的应用算法人员处理数据、选择算法的经验

3.3 Causal Relationship Detection

Distinguishes correlation, association, and causation, and outlines statistical tests (Pearson, Spearman, Kendall) and methods (chi‑square, Fisher exact, hypergeometric, DAGs, Bayesian networks, Granger causality) for identifying causal links.

5. Application Cases

Case study on hotel business conversion rate decline demonstrates the four‑step workflow: problem detection, cause identification via binary‑tree reports, qualitative analysis of controllable vs. uncontrollable factors, and causal chain extraction (e.g., age → new/old customer → core destination → membership → conversion).

6. Results

Efficiency: case analysis time reduced by 75%.

Quality: analysis reports are product‑oriented and explainable.

Business Impact: covered multiple business lines, uncovered hidden scenarios in ticketing.

External Sharing: presented at 2023 China Database Technology Conference; filed a patent.

7. Summary and Planning

Key takeaways: data quality is critical; adopt a question‑driven mindset; align attribution with business practice. Future plans include leveraging AIGC for causal detection, establishing AB‑experiment attribution standards, and expanding diverse application scenarios across the company.