| # UERP_Model Changelog |
| |
| ## svd_model_25m.pkl (current, recommended) |
| - Trained on MovieLens 25M ratings (25,000,095 ratings, 162,541 users, 59,047 movies) |
| - SVD, n_factors=50, n_epochs=20 |
| - RMSE: 0.7728 | MAE: 0.5832 (held-out 10% test split) |
| - Serving: min_rating_count=100 filter applied |
| |
| ## svd_model_v1.pkl (deprecated, kept for comparison) |
| - Trained on MovieLens-latest-small (100,836 ratings) |
| - RMSE: 0.8775 | MAE: 0.6742 |
| |
| ## ncf_model_v2.keras (Stage 4 — Neural Collaborative Filtering) |
| - Trained on MovieLens 25M, embedding_dim=64, 3-layer MLP (256-128-64) |
| - RMSE: 0.8013 | MAE: 0.6030 (original 0.5-5 scale) |
| - Comparison: underperforms SVD baseline (RMSE 0.7728) even after tuning attempt (32-dim -> 64-dim) |
| - Finding consistent with Rendle et al. (RecSys 2020), "Neural Collaborative Filtering vs. |
| Matrix Factorization Revisited" — well-tuned MF/dot-product often outperforms learned-MLP |
| similarity for pure rating-prediction tasks. |
| - Decision: SVD remains primary CF signal for MVP serving. NCF kept as a documented |
| experiment / potential weak-signal contributor for future Stage 6 ranking ensemble. |
