Explainability Practices in JD Retail Recommendation System

The article introduces the practice of explainability in JD Retail's recommendation system, outlining its definition and the three core aspects: ranking explainability, model explainability, and traffic explainability.

Ranking Explainability records the flow of items through recall, filtering, coarse ranking, fine ranking, and strategy stages, providing a foundation for diagnosing issues such as missing recommendations, similar item overload, scenario mismatches, purchase cycle problems, and ordering anomalies.

Model Explainability addresses both global and local interpretation. Global methods explain feature contributions across the entire dataset (e.g., SHAP values for tree models), while local methods explain individual predictions (e.g., Integrated Gradients for neural networks). The article shows how these techniques identify why certain SKUs receive high or low scores.

Traffic Explainability analyzes overall product flow, long‑tail effects, and user behavior root causes, offering funnel visualizations and stage‑level SKU metrics.

The system architecture faces challenges of massive data volume (petabyte‑scale), governance inconsistencies, and diverse scenarios. JD leverages Flink for real‑time data ingestion and ClickHouse for multidimensional analysis, building dedicated pipelines for ranking and traffic explainability.

Key components include user profiling (demographic, value, preference), behavior profiling (macro actions like clicks and purchases, plus fine‑grained signals), and product profiling (category, brand, and attribute information). These profiles feed both recall and explainability modules.

Model explainability techniques highlighted are Tree SHAP for tree‑based models and Integrated Gradients for deep networks, with visual examples of feature importance distributions and SKU‑level comparisons.

Operational use cases demonstrate how explainability helps resolve bad cases, compare SKU performance, and understand user‑specific exposure decisions.

The article concludes with acknowledgments to JD Retail technical experts and the data team.