Data flow · Churn prediction

Leakage-free churn pipeline

Leakage-free churn pipeline A synthetic user-activity stream is split into chronological snapshots, a leakage guard blocks the late-joined post-snapshot tail from reaching the as-of feature build, and the trained LightGBM and logistic-regression models are judged by lift at the top decile and then re-stratified on the at-risk subgroup. 01INGEST 02WINDOW 03GUARD 04AS-OF 05MODEL 06STRATIFY SYNTHETICSOURCE AS-OFFEATURES MODELTRAINING EVALUATIONSTRATIFY AS-OF WINDOWS FORBIDDEN SRC Raw Events users.csv · activity.csv DB SRC Post-Snapshot Tail future · late-joined DB FEA Snapshot Split train · val · test DB TB FEA Leakage Guard never sees the future DB FEA As-of Features prior 7/14/30d windows TB TB FEA Churn Label silent [snap,+30d] TB TB MOD Model Train LightGBM · LogReg TB FL EVA Lift@Top-Decile recall 0.544 · 5.44× FL FL EVA At-Risk Subset 1.75× · base 0.572 FL FL NODE PIPELINE STAGE FOCAL · GUARD OPTIONAL · LATE-JOINED DATA DB DATASET TB TABLE FL EXPORT LEFT CHIP = INPUT · RIGHT CHIP = OUTPUT FLOW STANDARD HANDOFF FOCAL HANDOFF GOVERNANCE TRIGGER FORBIDDEN · DROPPED