A group preference-based item similarity model: comparison of clustering techniques in ambient and context-aware recommender systems

Abstract

In the study of collaborative filtering, scholars and professionals alike have given much attention to user responses of the “one-class” type, feedback like online transactions or “likes”. Such behavior gauges have been integral to many ambient intelligent and context-aware recommendation systems, in which users are furnished with personalized lists of items according to their exhibited tastes. These one-class data, earlier studies have shown, are readily grasped by Bayesian personalized ranking, a pairwise preference assumption. Nevertheless, these works fail to make sufficient use of item similarity models using group preference. To improve performance, we argue in this paper, it is necessary to develop a model that yokes a User preference model to the Group Preference-based Similarity models (called UGPS). UCPG will produce a greater depth of interactions, we argue, because it takes on an entire set of items as opposed to the solitary item used previously. Moreover, a number of clustering methods have been put to work in group preference-based recommendation systems, but there is no consensus as to which offers superior accuracy. To gain clarity, we first built up a pair of versions of UGPS in order to assess the recommendation performances of different approaches to group generation: UGPS-1, which employed K-means, and UGPS-2, using K-NN—according to how efficiently they group their output. This comparison revealed that UGPS-1 tended to improve its recommendation performance as the number of groups and representative item sets grew. In contrast, UGPS-2 exhibited the opposite effect: recommendation performance declined as the number of groups and representative item sets expanded. Lastly, we consider how our UGPS system works with various sophisticated approaches on four real datasets, and demonstrate that UGPS produces more accurate recommendations.