International Journal of Computer Science Issues

More than a traditional journal...

Collaborative filtering systems offer customized recommendations to users by exploiting the interrelationships between users and items. Users are assessed for their similarity in tastes and items preferred by similar users are offered as recommendations. However scalability and scarcity of data are the two major bottlenecks to effective recommendations. With web based RS typically having users in order of millions, timely recommendations pose a major challenge. Sparsity of ratings data also affects the quality of suggestions. To alleviate these problems we propose a genetic programming approach to feature extraction by employing GP to convert from user-item space to user-feature preference space where the feature space is much smaller than the item space. The advantage of this approach lies in the reduction of sparse high dimensional preference information into a compact and dense low dimensional preference data. The features are constructed using GP and the individuals are evolved to generate the most discriminative set of features. We compare our approach to content based feature extraction approach and demonstrate the effectiveness of the GP approach in generating the optimal feature set.