Designing Specific Weighted Similarity Measures to Improve Collaborative Filtering Systems

The aim of collaborative filtering is to help users to find items that they should appreciate from huge catalogues. In that field, we can distinguish user-based from item-based approaches. The former is based on the notion of user neighbourhoods while the latter uses item neighbourhoods. The definition of similarity between users and items is a key problem in both approaches. While traditional similarity measures can be used, we will see in this paper that bespoke ones, that are tailored to type of data that is typically available (i.e. very sparse), tend to lead to better results. Extensive experiments are conducted on two publicly available datasets, called MovieLens and Netflix . Many similarity measures are compared. And we will show that using weighted similarity measures significantly improves the results of both user- and item-based approaches.