How pscredefine/rankshift works
RankShift is a research and recommender-system repository: it measures how recommendation models behave when exposure policy changes, then includes a two-stage retrieval-and-ranking implementation with an HTTP recommendation endpoint. The research workflow runs experiments over KuaiRand logs and records metrics; the serving workflow retrieves candidates, builds point-in-time features, ranks them, and returns recommendations. The README documents several experiment capabilities whose source was not sampled; those links are shown as documented relationships rather than claims of source inspection.
Main entry points: results Experiment results · 07_two_stage.py Two-stage experiment · pipeline.py Recommendation pipeline · features.py Feature builder
Components 18
Exposure research
- Signal probe
01_signal_probe.py - Exposure evaluation
02_exposure_bias.py - Calibration study
03_recalibrate.py - Off-policy evaluation
04_off_policy.py - Neural baselines
05_two_tower.py - Scale evaluation
06_scale_27k.py
Recommender runtime
- Two-stage experiment
07_two_stage.py - Feature contracts
recsys/schema.py - Feature builder
recsys/features.py - Candidate retrieval
recsys/models.py - Ranking models
recsys/models.py - Recommendation pipeline
recsys/pipeline.py - Versioned bundles
recsys/pipeline.py
HTTP serving
- Recommendation API
recsys/api.py
Other components
- Engineer
- KuaiRand logs
- Experiment results
results - API client
Connections 32
- Engineer → Signal probe: runs experiment
README.md - Engineer → Exposure evaluation: runs experiment
README.md - Engineer → Calibration study: runs experiment
README.md - Engineer → Off-policy evaluation: runs experiment
README.md - Engineer → Neural baselines: runs experiment
README.md - Engineer → Scale evaluation: runs experiment
README.md - Engineer → Two-stage experiment: runs experiment
README.md - KuaiRand logs → Signal probe: supplies logs
- KuaiRand logs → Exposure evaluation: supplies logs
02_exposure_bias.py - KuaiRand logs → Off-policy evaluation: supplies logs
04_off_policy.py - KuaiRand logs → Neural baselines: supplies logs
05_two_tower.py - KuaiRand logs → Two-stage experiment: supplies logs
07_two_stage.py - Signal probe → Experiment results: writes metrics
README.md - Exposure evaluation → Experiment results: writes metrics
02_exposure_bias.py - Calibration study → Experiment results: writes metrics
README.md - Off-policy evaluation → Experiment results: writes estimates
README.md - Neural baselines → Experiment results: writes metrics
README.md - Scale evaluation → Experiment results: writes metrics
README.md - Two-stage experiment → Experiment results: writes metrics
README.md - Two-stage experiment → Feature builder: builds features
07_two_stage.py - Two-stage experiment → Candidate retrieval: fits retriever
07_two_stage.py - Two-stage experiment → Ranking models: fits rankers
07_two_stage.py - Feature builder → Feature contracts: uses contracts
recsys/features.py - Candidate retrieval → Feature builder: uses encodings
recsys/models.py - Ranking models → Feature builder: uses encodings
recsys/models.py - API client → Recommendation API: requests recommendations
recsys/api.py - Recommendation API → Recommendation pipeline: loads and calls
recsys/api.py - Recommendation pipeline → Candidate retrieval: retrieves candidates
recsys/pipeline.py - Recommendation pipeline → Feature builder: builds request features
recsys/pipeline.py - Recommendation pipeline → Ranking models: scores candidates
recsys/pipeline.py - Recommendation API → Versioned bundles: loads bundle
recsys/api.py - Recommendation pipeline → Versioned bundles: saves or loads
recsys/pipeline.py