Book Recommendation using Collaborative Filtering (Book-Crossing)

'Readers who liked X also liked Y' is the oldest trick in recommendation — and still the backbone of how bookstores, libraries and reading apps suggest your next book. The idea is collaborative filtering: find items similar to the ones you liked, where similarity is learned from the rating behaviour of thousands of readers, not from genre tags. The Book-Crossing dataset is the public benchmark for this — 1.15 million ratings across 278,858 users and 271,379 books. This project builds item-based collaborative filtering on it properly: a user–item rating matrix, item–item cosine similarity, score aggregation, top-N ranking, and evaluation with Precision@5 and Recall@10. The interactive demo makes it tangible — pick books you liked and watch the evidence-cited recommendations appear.

1. The Book-Crossing ratings (1–10 scale) are loaded, cleaned and assembled into a user–item rating matrix.
2. Item–item cosine similarity is computed across the rating vectors, capturing which books are liked by the same readers.
3. For a user's liked books, candidate scores are aggregated as Σ similarity × rating and ranked to produce the top-N list, each with its evidence book.
4. A popularity baseline and a user-based CF variant are built for comparison on the same held-out ratings.
5. Evaluation on held-out ratings measures Precision@5 and Recall@10 for all three methods; the comparison is reported honestly.
6. In the web demo, your book picks are scored against the fitted similarities and the ranked, evidence-cited recommendations render instantly.

Tech stack:

• Python 3, Pandas, NumPy (matrix handling)
• scikit-learn (cosine similarity, metrics)
• Matplotlib (evaluation plots)
• Jupyter Notebook (analysis & evaluation)
• HTML5 + JavaScript (interactive recommender demo)
• Book-Crossing dataset (1.15M ratings)

Dataset & model details

• Dataset: Book-Crossing (Cai-Nicolas Ziegler) — 1.15 million ratings (1–10 scale) from 278,858 users on 271,379 books; the standard public benchmark for book recommendation.
• Task: Top-N recommendation; input = user's liked books, output = ranked book list with evidence citations.
• Model: Item-based collaborative filtering — item–item cosine similarity on the user–item matrix, score = Σ similarity × rating, top-N ranking. Compared against user-based CF and a popularity baseline.
• Metrics: Precision@5 ≈ 0.31, Recall@10 ≈ 0.42 (design targets for the built-to-order run) on held-out ratings, plus coverage over the catalogue. No figure is claimed as measured until the run is executed for the order.

• [Item-based collaborative filtering] Item–item cosine similarity computed on the user–item rating matrix; recommendations scored by Σ similarity × rating and ranked top-N — the canonical, explainable CF approach.
• [Explainable recommendations] Every suggestion cites the evidence book from your picks that drove it ('because you liked The Hobbit') — recommendations you can defend, not black-box scores.
• [Interactive recommender demo] Pick from 12 well-known books, get ranked recommendations instantly with match scores and evidence citations.
• [Real benchmark dataset] Book-Crossing — 1.15M ratings, 278,858 users, 271,379 books — the public standard for book recommendation research.
• [Ranking evaluation] Precision@5 ≈ 0.31 and Recall@10 ≈ 0.42 (design targets) measured on held-out ratings — the correct metrics for recommenders, not classification accuracy.
• [Baseline comparison] Item-based CF compared against user-based CF and a popularity baseline, so the report shows why the chosen method wins.
• [Full analysis notebook] Matrix construction, similarity computation, ranking, evaluation and baseline comparison in one reproducible Jupyter notebook.
• [Exported similarity model] Fitted item similarities saved, so the demo recommends with the real model instantly.

• Complete analysis Jupyter notebook (similarity, ranking, evaluation, baselines)
• Fitted item-similarity model files
• Interactive recommender web demo with explainable suggestions
• Precision/recall plots and baseline comparison figures
• Project report PDF (background, CF theory, methodology, evaluation, cold-start discussion)
• PPT presentation for final review
• Viva Q&A preparation document (user-based vs item-based CF, cosine similarity, sparsity, cold start, ranking metrics)

• Cold start: new users with no ratings and new books with no ratings get weak recommendations — the report discusses this fundamental CF limit honestly.
• The 1.15M-rating matrix is sparse; similarity quality depends on rating density, discussed in the report.
• Popularity bias: widely-rated books dominate suggestions; the report covers this and lists diversification as future scope.
• The demo uses a 12-book catalogue to illustrate the method; the shipped build runs on the full Book-Crossing matrix.
• All metric figures are design targets for the run, stated honestly — the report documents the actual achieved figures.