SCCF Parameter and Similarity Measure Optimization and Evaluation

Abstract

Neighborhood-based Collaborative Filtering (CF) is one of the most successful and widely used recommendation approaches; however, it suffers from major flaws especially under sparse environments. Traditional similarity measures used by neighborhood-based CF to find similar users or items are not suitable in sparse datasets. Sparse Subspace Clustering and common liking rate in CF (SCCF), a recently published research, proposed a tunable similarity measure oriented towards sparse datasets; however, its performance can be maximized and requires further analysis and investigation. In this paper, we propose and evaluate the performance of a new tuning mechanism, using the Mean Absolute Error (MAE) and F1-Measure metrics, in order to show that the SCCF similarity can be radically enhanced and thus increasing its overall efficiency. Moreover, the SCCF similarity measure was tested against several other measures, targeted especially at sparse datasets, and the results show how the effectiveness of a measure significantly varies with the dataset structure and properties, and that one measure cannot be considered as better than the other when compared with a small selection of other measures.