EXP-005 / question
Technical question
When does a hybrid recommender actually beat simple recommendation baselines?
EXP-005 / method
Method and workflow
- Generate deterministic synthetic movies and user ratings.
- Create train/test splits and user-item matrices.
- Build content-based recommendations with TF-IDF and cosine similarity.
- Build collaborative recommendations with corrected TruncatedSVD reconstruction.
- Compare hybrid blends across alpha values from 0.0 to 1.0.
- Report baselines and structured per-user recommendation outputs with short reasons.
EXP-005 / evidence
Evidence of work
Bayesian-average, popularity, average-rating, positive-count, and random baselines make model quality honest.
Hybrid weights are measured instead of guessed, with the current synthetic dataset favoring content-only blending.
Recommendation CSVs include rank, movie ID, title, genres, score, and reason fields.
EXP-005 / stack
Technical stack
EXP-005 / limitations
Limitations and honesty check
- The dataset is synthetic and should not be interpreted as real user preference behavior.
- The current best model can be a simple baseline, which is intentionally documented rather than hidden.
- Production recommenders would need real interaction logs, ranking experiments, cold-start handling, and online evaluation.
EXP-005 / next
Next improvements
- Add optional MovieLens support.
- Add item/user cold-start diagnostics.
- Add ranking diversity and novelty metrics.
- Add a small recommendation explorer UI.