Multi-Stage Ranking

Source: https://medium.com/better-ml/multi-stage-ranking-e0dacd81ac4 Author: Jaideep Ray (Sr. Staff ML Engineer)

Summary

Practical explanation of how production search and recommendation systems break ranking into sequential ML stages, each optimizing different objectives with different speed/quality trade-offs.

The Three Stages

Stage 1 — Retrieval

Goal: Maximize recall, minimize latency. Method: Gradient boosted decision trees with few high-quality features. Scale: Thousands of candidates → hundreds.

Stage 2 — Relevance Ranking

Goal: Optimize document/author quality and relevance. Method: Stacked ensembles using human-judged training data. Scale: Hundreds → tens of candidates.

Stage 3 — Click Ranking

Goal: Maximize engagement via learned user behavior. Method: Deep neural networks (Wide & Deep), abundant click data. Scale: 10-100 candidates → final ranked list.

Key Trade-offs

Benefits:

  • Performance optimization at each scale
  • Latency management per stage
  • Decoupled development — teams iterate independently

Drawbacks:

  • “Makes overall system hard to debug”
  • Cascading failures — errors in Stage 1 propagate
  • “Hard to maintain and iterate on” across stages

Engineering Considerations

  • Feature versioning prevents cascade failures
  • Data linting at each stage catches feature distribution drift
  • Fast offline testing enables rapid model iteration
  • A/B testing validates major architecture changes