Recommender system algorithms: A comparative analysis based on monotonicity

Recommender systems are designed in such a way that they sort through massive amounts of data so as to help users in finding their preferred items. Currently much research on recommender systems focus on improving the prediction or classification accuracy of the respective algorithms while behavioral aspects are often overlooked. In this paper we focus on a particular behavioral property called monotonicity which we believe every recommender system should satisfy but has received very little attention. Assume that a recommender system recommends a set R of N items (for example movies) based on some scoring pattern and o_a ∈ R has the highest score among items in R. Let us also assume that a particular user accepts the recommendation and consumes that item o_a in the next stage and thereby appends his/her profile with o_a. We again run the recommendation step with the appended profile to get a recommendation of N items, say R´. Monotonicity refers to the number of items of R {_a} (R minus {o_a}) that are retained in R´. In a top-k recommendation, monotonicity tries to measure the number of items continued to be recommended when a technique is utilized incrementally. In this work, in addition to monotonicity, we also consider two other popular measures called precison and recall to provide an experimental analysis of five most popular recommendation algorithms for evaluating the utility of recommendations.