Understanding and Handling Bad Cases in E-commerce Recommendation Systems

This article examines the phenomenon of "bad cases"—situations where users encounter unsatisfactory or erroneous recommendations and search results in e‑commerce platforms—and explains why they inevitably arise due to imperfect models, biased data, and system constraints.

It categorizes bad cases into functional bugs, experience issues, political‑correctness problems, and ambiguous cases, and discusses their relative severity and impact on product health.

The author presents a systematic approach to analyzing bad cases: collecting logs, measuring user behavior metrics (e.g., refresh rates, dwell time), and performing both online and offline investigations to identify root causes such as feature gaps, model shortcomings, or data noise.

Several real‑world examples are detailed, including a case where a user’s pull‑to‑refresh did not change recommendations, a mismatch between category distribution in "猜你喜欢" versus search, and lack of diversity in recommendation lists.

Based on these analyses, the article proposes mitigation strategies ranging from quick rule‑based interventions (e.g., filtering illegal content, mapping query synonyms) to more complex model‑centric solutions (feature engineering, sample cleaning, model fusion between NN and GBDT, listwise ranking for diversity).

An iterative closed‑loop process is outlined: collect bad cases, analyze data, apply rule or model changes, run A/B tests, and refine the system continuously, while also emphasizing the importance of monitoring, logging, and cross‑team collaboration.

Finally, the author highlights how systematic bad‑case handling deepens business understanding, improves long‑term metrics such as retention and LTV, and encourages algorithm engineers to balance model innovation with practical product priorities.