An improved item-based collaborative filtering using a modified Bhattacharyya coefficient and user–user similarity as weight

Item-based filtering technique is a collaborative filtering algorithm for recommendations. Correlation-based similarity measures such as cosine similarity, Pearson correlation, and its variants have inherent limitations on sparse datasets because items may not have enough ratings for predictions. In addition, traditional similarity measures mainly focus on the orientations of the rating vectors, not magnitude, and as a result two rating vectors with different magnitudes but oriented in the same direction, can be exactly similar. Another aspect is that on a set of items, similar users’ may have different rating pattern. In addition, to calculate the similarity between items, ratings of all co-rated users are considered; however, a judicious approach is to consider the similarity between users as a weight to find the similar neighbors of a target item. To mitigate these issues, a modified Bhattacharyya coefficient is proposed in this paper. The proposed similarity measure is used to calculate user–user similarity, which in turn is used as a weight in item-based collaborative filtering. The experimental analysis on the collected MovieLens datasets shows a significant improvement of item-based collaborative filtering, when user–user similarity calculated by the proposed modified similarity measure is used as a weight.