A Hybrid Approach to Recommender Systems based on Matrix Factorizations
Due to the huge amount of information available online, the need of personalization and filtering systems is growing permanently. Recommendation systems constitute a specific type of information filtering technique that attempt to present items according to the interest expressed by a user. Commonly online recommender are employed for e-commerce applications or customer adapted websites. In general, there exist two basic types of recommendation techniques, namely content-based filtering and collaborative filtering. Whereas content-based filtering methods examine items previously favored by the actual user, collaborative filtering computes recommendations based on the information about similar items or users. In our work we combine both techniques into a hybrid approach, where supplementary content features are employed to improve the accuracy of collaborative filtering. For the development of our hybrid recommender we utilized the well-known MovieLens rating data as well as the IMDB online movie archive. The content information retrieved from IMDB is converted into a notation that is useable for our hybrid approach. Rating and content data are both normalized separately, before the combined information is utilized by our recommendation algorithm. In order to reduce the computational effort of our hybrid model, we furthermore factorize the extended rating matrix by means of singular value decomposition. A prototype system of our novel hybrid recommender was implemented in MATLAB programming language. By means of various experiments, we could demonstrate that the extracted content features are beneficial to our underlying rating prediction algorithm. In addition, we discover a way to reveal latent feature relations, which can be used to generate more individual and accurate recommendations.