Fix pipeline data leakage, merge fan-out, and repo hygiene

[skylarliu1]

What this fixes

1. Data leakage in every supervised notebook (primary)

All 7 classifiers merged the unsupervised artifacts (dbscan_df.csv, gmm_df.csv,
kmeans_clusters.csv, renttherunway_lda_topics.csv) into df before
train_test_split(). Those artifacts were produced by models (DBSCAN, GMM,
K-means, LDA) fit on the entire RTR dataset, so test-row cluster labels and
topic vectors were informed by the test set.

Fix: new src/split.py defines one frozen split (row-level, stratified on
fit_label, test_size=0.20, random_state=42). Every Phase 1 notebook now fits
its scaler / model / PCA / model-selection sweep on the training partition only
and merely transforms every row:

model	how test rows are assigned
GMM, K-means	predict / predict_proba
DBSCAN (no predict)	nearest fitted core sample within eps, else -1
LDA	Dictionary + LdaModel built from train reviews; test reviews scored with get_document_topics (inference only). Train-only coherence sweep now picks k=9 (was k=11 on the full corpus).

Every Phase 2 notebook takes its train/test partition from df["split"] instead of
a fresh train_test_split.

2. Row-count fan-out bug

dbscan_df.csv was transaction-grained but keyed only on user_id (~70k duplicate
keys). df.merge(dbscan, on="user_id") fanned 192,544 → 1,156,140 rows, which
is where the "test set of 266,286 samples" came from. The other three merges added
smaller fan-out from 266 genuine duplicate (user_id, item_id) transactions in
renttherunway_clean.csv.

Fix: each artifact is de-duplicated on its merge key at the source
(dbscan_df.csv → one row per user_id; the others → unique (user_id, item_id[, fit, fit_label])),
and each supervised notebook re-asserts len(df) == 192_544 after the merges.
Clean 20% test set is now 38,509 rows.

3. Repo hygiene

• New .gitignore (venvfit/, .DS_Store, __pycache__/, Data/, Models/).
• New requirements.txt (pinned to the project venv).
• git rm --cached Data/Raw/, Data/Processed/ (regenerable via
  download_data.py / data_cleaning.py / the notebooks) and Models/
  (regenerable by re-running Phase 2; bagging_simple_classifier.joblib alone is
  160 MB and was the repo's only Git-LFS entry — its .gitattributes rule is
  removed). Local files are kept; this is a normal commit, not a history
  rewrite. venvfit/ stays in history as-is per the brief.
• README.md: added a "Reproducing the Pipeline" run-order section.
• Fixed a pre-existing figure/data path bug in DBSCAN, K-means, and
  Model_Comparison (../ → ../../, exposed by the earlier repo restructure).

Impact on results

Removing the leakage (and the fan-out) collapses every model to near the
majority-class base rate:

model	leaky acc / macro-F1	fixed acc / macro-F1
CART	0.73 / 0.67	0.50 / 0.38
Random Forest	0.76 / 0.70	0.53 / 0.40
Bagging	0.77 / 0.72	0.52 / 0.40
KNN	0.91 / 0.87	0.69 / 0.34
Kernel SVM	0.85 / 0.78	0.52 / 0.40
Linear SVM	0.68 / 0.39	0.74 / 0.29
Naive Bayes	0.46 / 0.41	0.42 / 0.35

Model_Comparison now shows the rich RTR feature set is worse than the simple
3-feature (weight/age/height) baseline for every model (macro-F1 Δ −0.23 to
−0.40). Its "KNN on rich features is the strongest overall performer" conclusion no
longer holds.

Follow-ups for reviewers (not done here)

1. Write-up prose. Every model's markdown write-up + the LDA and Model_Comparison
   write-ups have a ⚠️ banner with the fixed headline numbers, but the narrative
   paragraphs, per-class tables, and confusion-count callouts still describe the
   leaky results and need a human rewrite.
2. Same user in train and test. Body measurements are constant per user_id, so
   a user with rows on both sides still leaks (weaker form). Kept a plain row-level
   split here to stay comparable; consider GroupShuffleSplit on user_id.
3. predict vs argmax(predict_proba). Model_Comparison scores the RTR models
   from the saved probability files, which disagrees with each notebook's own
   predict() for the class_weight="balanced" models (e.g. Kernel SVM: 0.52 vs
   0.74 acc). Pre-existing; worth reconciling.
4. Data/ and Models/ are no longer in git — run src/download_data.py →
   src/data_cleaning.py → the notebooks (order in the README) to regenerate.

How it was verified

src/split.py → 154,035 train / 38,509 test (exactly 20%, stratified). All 11
notebooks (4 unsupervised + 7 supervised) + Model_Comparison were executed
end-to-end on venvfit with a venvfit kernel; each supervised notebook prints
train rows: 154,035 / test rows: 38,509 and its merge-fan-out assertion passes.

🤖 Generated with Claude Code