A Comparative Study on Prediction Approaches of Item-Based Collaborative Filtering in Neighborhood-Based Recommendations

With the growing nature of data over the internet, item-based collaborative filtering has become a promising method in the recommendation. The two-step process of item-based collaborative filtering, i.e., computation of similarity among items, and rating prediction using similar items are utilized in recommendation. However, the quality of recommendations after following these steps degrade in sparse datasets. Traditionally, in item-based collaborative filtering, several similarity measures have used to find top-k similar items, and prediction approaches are utilized for rating prediction, and then a top-n list of recommended items is generated. Plenty of work has been done to increase the performance of collaborative filtering using the combination of new/modified similarity measures and the traditional prediction approach. But, traditional prediction approaches also give future scope for improvement in the recommendation system. Therefore, the objective of the paper is to serve a comparative study on conventional prediction approaches for existing best similarity measures. The performance of different prediction approaches is tested with MovieLens datasets using various accuracy metrics.