Beyond k-NN: combining cluster analysis and classification for recommender systems

Recommendation systems have a wide application in e-business and have been successful in guiding users in their online purchases. The use of data mining techniques, to aid recommendation systems in their goal to learn the correct user profiles, is an active area of research. In most recent works, recommendations are obtained by applying a supervised learning method, notably the k-nearest neighbour (k-NN) algorithm. However, classification algorithms require a class label, and in many applications, such labels are not available, leading to extensive domain expert labelling. In addition, recommendation systems suffer from a data sparsity problem, i.e. the number of items purchased by a customer is typically a small subset of all \u109vailable products. One solution to overcome the labelling and data sparsity problems is to apply cluster analysis techniques prior to classification. Cluster analysis allows one to learn the natural groupings, i.e. similar customer profiles. In this paper, we s tudy the value of applying cluster analysis techniques to customer ratings prior to applying classification models. Our HCC-Learn framework combines content-based analysis in the cluster analysis stage, with collaborative filtering in the recommending stage. Our experimental results show the value of combining cluster analysis and classification against two real-world data sets.Peer reviewed: YesNRC publication: Ye