Book Recommendation using Hybrid Filtering

Production recommenders are rarely pure collaborative filtering or pure content-based — they are hybrids, because each pure method has a blind spot the other covers. CF is strong where ratings are dense but helpless for a new user with two ratings; content-based methods survive cold start but can only recommend more-of-the-same and never surprise. Students typically implement the two as separate assignments and never see them combined: how two scores on different scales are normalized, how the blend weight is chosen by validation instead of gut feel, and how the system degrades gracefully when one signal is missing. The Goodbooks-10k corpus — 10,000 books, 6M ratings from 53,424 readers, plus genre and shelf-tag metadata — is the standard public book-rating dataset. This project builds both signals on it, sweeps the blend weight on a validation split, and ships a demo where lambda is a live slider, so the hybrid stops being an equation and becomes something you can feel.

1. Goodbooks-10k ratings.csv (1-5 stars), books.csv metadata and book_tags.csv shelf tags are joined; a train/validation/test split is fixed by seed.
2. The CF signal: user-based kNN with cosine similarity over mean-centered ratings predicts scores for books the user has not rated.
3. The content signal: TF-IDF over genres + top shelf tags, cosine similarity against the user's highly-rated books, averaged into one score per candidate.
4. Both signals are min-max normalized per user onto [0,1] and blended with lambda; the validation sweep picks the lambda maximizing nDCG@10 (expected near 0.6).
5. At serving time the demo computes both scores for candidates, blends at the current lambda, and renders the top 5 with per-signal contribution bars.
6. The evaluation notebook reports RMSE, Precision@10, nDCG@10 and catalog coverage for CF-only, content-only and hybrid side by side.

Tech stack:

• Python 3.10, scikit-surprise / scikit-learn (user-based kNN)
• scikit-learn TfidfVectorizer (genre + shelf-tag content vectors)
• pandas, NumPy (joins, normalization, lambda sweep)
• Matplotlib (lambda-sweep curve, metric comparison charts)
• Flask demo app (rating, comparison, blend-tuner and evaluation views)
• Goodbooks-10k dataset (10,000 books, 6M ratings, 53,424 users)

• [User-based kNN collaborative signal] Cosine similarity over mean-centered Goodbooks-10k ratings finds nearest-neighbor readers and predicts scores for unseen books.
• [TF-IDF content signal] Genres plus top user shelf-tags per book become TF-IDF vectors; similarity to the user's highly-rated books gives the content score.
• [Validated lambda blend] Both signals are min-max normalized per user and combined as lambda x CF + (1-lambda) x content, with lambda swept in [0,1] on validation maximizing nDCG@10.
• [Live blend tuner] The demo binds lambda to a slider that re-ranks the top 5 instantly, showing how low lambda favors safe genre matches and high lambda favors crowd wisdom.
• [Three-way method comparison] Flip between Hybrid, CF-only and Content-only rankings to see exactly what each signal contributes — and what each loses alone.
• [Per-signal contribution bars] Every hybrid recommendation shows its CF and content component scores, so the blend is auditable, not magic.
• [Cold-start fallback rules] With fewer than 3 ratings lambda is forced toward 0 (content leads); the demo's evaluation view documents the fallback ladder explicitly.

• Complete source code (kNN CF, TF-IDF content model, normalization, lambda sweep, Flask demo)
• Jupyter notebooks (buyer-run procedure: train both signals, sweep lambda, evaluate three-way)
• Evaluation outputs: lambda-sweep curve and metric comparison tables from your build
• Project report PDF (background, hybrid architectures, methodology, evaluation, cold-start analysis)
• PPT presentation for final review
• Viva Q&A preparation document (kNN, TF-IDF, score normalization, nDCG, hybrid design)
• Setup guide (environment, Goodbooks-10k download, running notebooks and demo)