Local Collaborative Filtering Optimization

Multi-objective optimization of Bicluster-based Collaborative Filtering variants for recommender systems

Investigation on how different variants of Local Neighborhood-based Collaborative Filtering (LNBCF) perform when optimizing multiple competing objectives in recommender systems.

The work specifically focuses on Bicluster-based Collaborative Filtering (BBCF) and proposes several modifications to try to enhance its performance. This research was presented as my Master's Thesis in order to attain the tittle of Master of Science in Electrical Engineering on the area of Computer Engineering at University of Campinas, Brazil.

Key Contributions

Comprehensive analysis of LNBCF variants using multi-objective optimization framework
Novel modifications to BBCF including:
- BinaPs Mining Algorithm (BMA) for pattern discovery
- Compound Weight Frequency (CWF) for pattern selection
- Cosine Similarity (CS) for rating prediction
- User-based Mode (UB) adaptation
Detailed evaluation showing trade-offs between prediction accuracy, coverage, and computational efficiency

Technical Details

The research employs the MovieLens 100K dataset and evaluates performance using multiple metrics:

Mean Absolute Error (MAE)
Precision@20
Recall@20
Normalized Discounted Cumulative Gain (nDCG@20)
Prediction Coverage
Total Runtime

Key Findings

No single LNBCF variant dominates across all objectives
Different variants excel in specific scenarios:
- User-based variants perform better in Recall@20 tasks
- Item-based variants excel in MAE optimization
- SVD variants show superior Precision@20 performance
Hyperparameter sensitivity varies significantly between variants and objectives

Implementation

The project is implemented in Python using: