Using Mutual Influence to Improve Recommendations

In this work we show how items in recommender systems mutually influence each other's utility and how it can be explored to improve recommendations. The way we model mutual influence is cheap and can be computed without requiring any source of content information about either items or users. We propose an algorithm that considers mutual influence to generate recommendations and analyse it over different recommendation datasets. We compare our algorithm with the Top - N selection algorithm and obtain gains up to 17% in the utility of recommendations without affecting their diversity. We also analyse the scalability of our algorithm and show that it is as applicable for real-world recommender systems as Top - N.