ADAditya Dandiainadityand.hashnode.dev·Apr 27 · 5 min readEpisode 17: Phase 6: Final Evaluation & Series FinaleAct 1 — The Full Picture Five phases in, eight models trained, one question left: which model do you actually deploy, and can it explain its decisions? The full comparison table tells the first part o00
ADAditya Dandiainadityand.hashnode.dev·Apr 26 · 3 min readEpisode 16: Phase 5: LightGBM Didn't Win the Way I Expected. It Won Differently.Act 1 — The Setup Phase 4 ended with LR + SMOTE as the model to beat. Churn recall of 0.80. ROC-AUC of 0.839. A logistic regression — the simplest model in the toolkit — holding the top spot after fou00
ADAditya Dandiainadityand.hashnode.dev·Apr 26 · 3 min readEpisode 15 — Phase 4: Class ImbalanceAct 1 — The Problem With Getting 81% Right Phase 3 ended with an uncomfortable truth. Logistic Regression scored 81% accuracy and an ROC-AUC of 0.842. On paper, that looks good. But it was missing 44%00
ADAditya Dandiainadityand.hashnode.dev·Apr 26 · 6 min readEpisode 14: Phase 3: Baseline ModelsTitle: I Ran Three Baseline Models on Churn Data. The Simplest One Won. Act 1 — The Setup Phase 2 left me with a clean preprocessing pipeline. A ColumnTransformer that handles three feature types, a 00
ADAditya Dandiainadityand.hashnode.dev·Apr 26 · 3 min readEpisode 13: Phase 2: The Preprocessing PipelineAct 1 — The Problem With Raw Data Raw data is a mess. Strings where there should be numbers. Categories with three possible values mixed alongside categories with seven. Numerical features on complete00