Book Recommender Systems — Full Taxonomy
Complete recommender system taxonomy on BookCrossing (1.1M ratings): User-CF, Item-CF, SVD/NMF/ALS, Content-Based, Hybrid, NCF, AutoRec, GRU4Rec. User-CF RMSE 1.6645, P@10 0.6629, R@10 0.6910.
BookCrossing: 1,149,780 ratings, 271,360 books, 278,858 users
Full taxonomy: User-CF → Item-CF → MF → Content-Based → Hybrid → NCF/AutoRec/GRU4Rec
Comprehensive implementation of all major recommender paradigms on the Book-Crossing dataset.
Dataset
- ▸Raw: 271,360 books, 278,858 users, 1,149,780 ratings
- ▸Filtered (explicit ≥ 1): 118,699 ratings, 7,027 users, 9,438 books
- ▸Rating scale: 1–10 (explicit) + implicit feedback (page views)
Full Taxonomy
Recommender Systems
├── 1. Collaborative Filtering
│ ├── User-based CF (cosine similarity, K=20)
│ ├── Item-based CF
│ └── Matrix Factorization: SVD / NMF / ALS
├── 2. Content-Based Filtering
│ └── TF-IDF on book metadata (title, author, genre)
├── 3. Hybrid
│ ├── Weighted combination (CF + CB)
│ └── Switching (CB for cold-start users)
└── 4. Deep Learning
├── NCF (Neural Collaborative Filtering)
├── AutoRec (Autoencoder CF)
└── GRU4Rec (Session-based sequential)
Key Results (User-CF)
| Metric | Value |
|---|---|
| RMSE | 1.6645 |
| Precision@10 | 0.6629 |
| Recall@10 | 0.6910 |
Cold-Start Handling New users (<5 ratings) fall back to Content-Based or popularity-based recommendations. Hybrid switching strategy avoids CF failure on cold-start.
Key Insight GRU4Rec (session-based) is the most practically valuable architecture — it handles anonymous users and captures short-term intent without needing a user history, which is the real-world default for most e-commerce sessions.