mratsim/Apartment-Interest-Prediction

Implements custom feature engineering pipelines with caching via `shelve` to overcome Scikit-Learn's limitations in handling non-leaky transformations across folds, combining XGBoost/LightGBM with HDBSCAN-based geo-clustering, TF-IDF+TruncatedSVD text decomposition, and cyclical time encoding via trigonometric projection. Custom cross-validation framework decouples feature transformations from fold computation, enabling efficient reuse of expensive NLP operations (sentiment analysis, metro vocabulary extraction) and categorical encodings across train/validation/test splits without data leakage.

No commits in the last 6 months.