One mode three-way overlapping cluster analysis (Q2430233)

A model is proposed for three-way similarities description (prediction) in terms of overlapping clusters. In this model, the similarity \(s_{ijk}\) between the objects \(i\), \(j\) and \(k\) is approximated by \[ \hat s_{ijk}=\sum_{r=1}^R w_r p_{ir}p_{jr}p_{kr}+C, \] where \(R\) is the number of clusters, \(p_{ir}=1\) if the \(i\)-th object belongs to the \(r\)-th cluster and 0 otherwise, \(w_r\) is the weight of the \(r\)-th cluster, and \(C\) is some constant. An iterative procedure is proposed for fitting this model, which is an extension of the MAPCLUS algorithm to deal with three-way similarities. Results of simulations and application to the analysis of customer preferences are presented.