Study and Implementation of Collaborative Filtering Algorithms for the Intelligent Generation of Recommendations under Cold Start Problem

Recommender Systems are software tools and techniques that aim to help users choose between a wide range of items. This diploma thesis contains an overview of the various approaches in the field of recommender systems. Emphasis was given on the collaborative filtering approach, since it is considered to be one of the most successful ones, and on the Cold-Start problem, that is a challenge for the generation of successful recommendations.

Hierarchical Itemspace Rank (HIR) is the algorithm that was chosen to be implemented. HIR tackles successfully the cold-start problem. It exploits the innate hierarchical structure of the itemspace to face the sparsity and cold-start problems and to generate qualitative recommendations. Except for the algorithm itself, a demo was implemented. This demo uses HIR to recommend movies to users, using a special for this purpose dataset. The recommendations are based on the users’ ratings on some movies.

The algorithm was implemented in Java using LensKit software. LensKit is a complete system for the implementation, comparison, experimental evaluation and research on recommender algorithms. It was chosen due to its features that enable the implementation of algorithms and the handling of the required data.

This implementation can be part of a new module in LensKit that would be specialized in facing cold-start problem.

AREA: Recommender Systems

KEYWORDS: Collaborative Filtering, Cold-Start Problem, HIR, LensKit