Interpretation cheat sheet (recsys-eval)¶
This page explains Interpretation cheat sheet (recsys-eval) and how it fits into the RecSys suite.
Before trusting any metric¶
- Validate schemas (extra/missing fields can break joins): Data contracts: what inputs look like
- Check join integrity:
- low match rate usually means broken instrumentation, not a “bad model”
- fix logging before debating metric moves
- Look for SRM warnings in experiments:
- SRM often indicates broken bucketing or assignment logging
- do not ship based on experiment results with SRM you can’t explain
If the primary KPI moved¶
Ask “is the move real, safe, and attributable?” in this order:
- Real: enough samples, stable joins, no obvious data anomalies.
- Safe: guardrails hold (latency, errors, empty recs, diversity constraints).
- Attributable: change is consistent across slices you care about.
Common “this looks wrong” signals¶
- KPI jumps by an impossible amount (often join issues or double-counting).
- Slice results contradict global results (often logging/slicing mismatch).
- High variance and no clear direction (often not enough traffic).
Read next¶
- Interpreting results: Interpreting results: how to go from report to decision
- Runbooks (common failure modes): Runbooks: operating recsys-eval
- Troubleshooting: Troubleshooting: symptom -> cause -> fix