How rajdeepofgithub/criteo-uplift works
This repository is a reproducible causal-analysis pipeline for the CRITEO-UPLIFT advertising benchmark. The user-facing workflow acquires and verifies data, ingests it, freezes a stratified split, audits the data, estimates the intention-to-treat effect, trains and compares fixed-budget targeting policies on validation, freezes the selected analysis, and opens a sealed test once. The graph retains leakage control, preprocessing, policy evaluation, and bootstrap uncertainty as internal layers. Several later-stage implementations are unsampled; their presence and purpose are based on the README, and those links cite it.
Main entry points: run_stage45_validation.py Validation modeling · run_stage2_split_audit_eda.py Split audit EDA · run_stage3_ate.py Assignment effect · leakage.py Feature whitelist
Components 21
Data preparation
- Acquire stage
scripts/run_stage1.py - Verified download
src/uplift_project/acquire.py - Parquet ingestion
src/uplift_project/ingest.py - Processed data
Study analysis
- Split audit EDA
scripts/run_stage2_split_audit_eda.py - Deterministic split
src/uplift_project/split.py - Split manifest
artifacts/manifests/split_manifest.json - Feature whitelist
src/uplift_project/leakage.py - Assignment effect
scripts/run_stage3_ate.py - ITT estimator
src/uplift_project/ate.py - Effect report
reports/ate/ate.json
Policy modeling
- Validation modeling
scripts/run_stage45_validation.py - Logistic preprocessing
src/uplift_project/preprocess.py - Budgeted targeting
src/uplift_project/policy.py - LightGBM selection
scripts/run_stage5b_lightgbm_v02.py
Evaluation and release
- Paired bootstrap
src/uplift_project/bootstrap.py - Freeze stage
scripts/run_stage6_freeze.py - Sealed test
scripts/run_stage6_sealed_test.py - Test report
reports/test/test.json
Other components
- Researcher
- Criteo benchmark
Connections 25
- Researcher → Acquire stage: runs pipeline
README.md - Acquire stage → Verified download: uses
scripts/run_stage1.py - Acquire stage → Parquet ingestion: uses
scripts/run_stage1.py - Verified download → Criteo benchmark: downloads
README.md - Parquet ingestion → Processed data: writes
README.md - Split audit EDA → Deterministic split: assigns split
scripts/run_stage2_split_audit_eda.py - Split audit EDA → Split manifest: writes
README.md - Split audit EDA → Feature whitelist: uses
scripts/run_stage2_split_audit_eda.py - Split audit EDA → Processed data: reads
scripts/run_stage2_split_audit_eda.py - Split audit EDA → Effect report: produces audit outputs
README.md - Assignment effect → ITT estimator: estimates effect
scripts/run_stage3_ate.py - Assignment effect → Feature whitelist: uses
scripts/run_stage3_ate.py - Assignment effect → Effect report: writes
README.md - Validation modeling → Logistic preprocessing: uses
scripts/run_stage45_validation.py - Validation modeling → Feature whitelist: guards features
scripts/run_stage45_validation.py - Validation modeling → Budgeted targeting: ranks policies
scripts/run_stage45_validation.py - Validation modeling → Paired bootstrap: estimates uncertainty
scripts/run_stage45_validation.py - Validation modeling → Processed data: loads split data
scripts/run_stage45_validation.py - LightGBM selection → Feature whitelist: guards features
scripts/run_stage5b_lightgbm_v02.py - LightGBM selection → Budgeted targeting: selects policies
scripts/run_stage5b_lightgbm_v02.py - LightGBM selection → Paired bootstrap: compares policies
scripts/run_stage5b_lightgbm_v02.py - Validation modeling → Freeze stage: informs selection
README.md - Freeze stage → Sealed test: freezes before
README.md - Sealed test → Test report: writes
README.md - Assignment effect → Processed data: analyzes
README.md