Analysis of similarity measures in user-item subgroup based collaborative filtering via genetic algorithm

Abstract

Recommender systems based on collaborative filtering (CF) analyze the mutual interests of similar users to predict the ratings of items for the active user (the user whom the prediction is for). This is done in two ways: (1) finding similarity between all users who share the same rating patterns with the active user (2) finding similarity between all pairs of items of different users by building an item-item matrix and thereafter inferring the tastes of the active user. We believe that considering the entire set of items is irrelevant in the prediction process and propose a user-item subgroup based local least squares CF technique that considers only a subset of highly correlated items based on set of similar users. An evolutionary algorithm framework is used to discover the subset of highly correlated items and a local least squares method is used to impute the missing ratings by analyzing the highly correlated user-item subgroup. As far as the experimental setup is concerned, three different similarity measures are used in the implementation of our proposed algorithm and a comparison is made with the traditional local least squares approach as well as with a state-of-the-art CF algorithm. Benchmark datasets like MovieLens are used for the experimental results.