Improving the performance of web service recommenders using semantic similarity

TLDR: The experiments described in the paper show that the focus is on reducing the problems arising from data sparsity, one of the main difficulties for CF algorithms, has a significantly positive impact on prediction accuracy, particularly when the user-item matrix is sparse.

Abstract: This paper addresses issues related to recommending Semantic Web Services (SWS) using collaborative filtering (CF). The focus is on reducing the problems arising from data sparsity, one of the main difficulties for CF algorithms. Two CF algorithms are presented and discussed: a memory-based algorithm, using the k-NN method, and a model-based algorithm, using the k-means method. In both algorithms, similarity between users is computed using the Pearson Correlation Coefficient (PCC). One of the limitations of using the PCC in this context is that in those instances where users have not rated items in common it is not possible to compute their similarity. In addition, when the number of common items that were rated is low, the reliability of the computed similarity degree may also be low. To overcome these limitations, the presented algorithms compute the similarity between two users taking into account services that both users accessed and also semantically similar services. Likewise, to predict the rating for a not yet accessed target service, the algorithms consider the ratings that neighbor users assigned to the target service, as is normally the case, while also considering the ratings assigned to services that are semantically similar to the target service. The experiments described in the paper show that this approach has a significantly positive impact on prediction accuracy, particularly when the user-item matrix is sparse.