Recommender System for Books

Dec 13, 2024

• Python
• pandas
• NumPy
• scikit-learn

• Best model RMSE: 0.82 (held-out ratings)
• Compared ALS, user-kNN, item-kNN, and metadata-enhanced hybrid CF
• Grade: 5.25 / 6

TL;DR

• Tackled the book recommendation problem on an extremely sparse user–item ratings dataset.
• Implemented ALS, user-based CF, item-based CF (with metadata), and a hybrid CF approach.
• Achieved best performance with the hybrid CF + metadata model: RMSE 0.82.
• Showed that metadata can add valuable information to improve recommendations.

At a glance

• Role: ML engineer (team of 2)
• Timeline: Oct–Dec 2024 (7 weeks)
• Context: EPFL Distributed Information Systems course project
• Users/Stakeholders: Students, researchers, and book recommendation engines
• Scope: Modeling, Hyperparameter tuning, Metadata integration, Evaluation

Problem

Recommender systems must predict user preferences despite sparse data: most users rate only one or two items, and new books suffer from cold-start issues. Traditional collaborative filtering (CF) struggles in such settings. The challenge: can we build a scalable recommender that performs well under sparsity, while still personalizing recommendations?

Solution overview

We implemented and compared multiple CF models:

1. ALS matrix factorization — baseline latent factor model.
2. User-based kNN CF — leverages user similarity for predictions.
3. Item-based kNN CF with metadata — enriches similarity with book metadata (subjects, summaries, language).
4. Hybrid CF — averages predictions from user- and item-based CF to balance their strengths.

Architecture

• Ratings matrix → split into train/val/test.
• ALS learns latent factors.
• kNN finds nearest neighbors (users or items) via cosine similarity.
• Metadata (ISBN-based features clustered with Sentence-BERT + k-means) enriches item similarity.
• Hybrid CF combines user- and item-CF predictions.

Data

• Dataset: Sparse user–book ratings matrix (explicit feedback).
• Sparsity: ~99.97% missing values (~100k ratings across ~19k users × ~15k books).
• Metadata: Subjects, summaries, and language from ISBNs. Encoded as categorical features and via clustering on Sentence-BERT embeddings.
• Splits: Train/val/test with only users and items appearing in train included in validation/test.

Method

• ALS: Latent factor decomposition with tuned factors (k=50), regularization (λ=0.8), 20 iterations.
• User-based CF: Cosine similarity over user rating vectors; neighborhood size tuned via grid search.
• Item-based CF: Cosine similarity over item rating vectors; metadata clusters added to similarity computation.
• Hybrid CF: Weighted average of user- and item-based predictions; optimized over validation set.

Experiments & Results

Benchmarks

A value of k = 5 was chosen for kNN models based on validation RMSEs.

Evaluation protocol.

• Metric: RMSE on held-out ratings.
• Hyperparameter tuning:
  • ALS: latent factors, regularization, number of parameters.
  • Collaborative filtering: kNN
  • Metadata clusters: k-means clustering.

Analysis

• ALS struggled with sparsity (RMSE > 1.1).
• User-based CF outperformed item-based CF slightly.
• Metadata improved item-based similarity but only marginally.
• Hybrid CF consistently yielded the best balance, with RMSE 0.824.

Impact

• Showed that classic CF methods remain strong baselines, even under severe sparsity.
• Demonstrated that metadata integration enhances methods, though with small improvements.
• Provided a scalable hybrid method with the strongest performance for this dataset.

What I learned

• Importance of hyperparameter tuning for CF stability.
• How sparsity impacts user- and item-based methods.
• Metadata helps but doesn’t always translate into large RMSE gains.
• Hybrid models are often more robust than single approaches.

Future Work

• Extend hybrid approach with weighted ensembles instead of simple averages.
• Incorporate implicit feedback (clicks, views) in addition to explicit ratings.
• Experiment with neural recommenders (Autoencoders, LightGCN) on this dataset.
• Explore real-world deployment with scalability and latency constraints.

References

• Scikit-learn NearestNeighbors documentation.
• Reimers & Gurevych (2019): Sentence-BERT.
• Rafailov et al. (2023): Hybrid CF strategies.