Optimal User Categorization from a Hierarchical Clustering Tree for Recommendation

Abstract

Recommender system (RS) and clustering are two main types of data mining techniques that have wide applications. An RS helps users to acquire useful online resources efficiently and effectively, whereas clustering groups similar objects together and separates dissimilar objects as much as possible. Recently, in an increasing number of studies, similar users/items are grouped before recommendation to improve the recommendation quality. Following this routine, we group similar users from a binary cluster tree formed using hierarchical clustering, for which no user-determined number of clusters is required. To extract optimal clusters, we incorporate the formation order of clusters, and propose two cluster quality measures based on lifetime and variance. The first measure favors stable clusters and the second measure tends to group users whose ratings have low variance. Extensive tests on public datasets demonstrate the superiority of the proposed method.