A Frequency-based Stochastic Blockmodel

Abstract: We propose a frequency-based infinite relational model (FIRM), which takesinto account the frequency of relation whereas stochastic blockmodels ignore frequency.We also derive a variational inference method for the FIRM to apply to a large dataset.Experimental results show that the FIRM gives better clustering results than a stochasticblockmodel on a dataset which has the frequency of relation.key words : stochastic blockmodels, Dirichlet process, variational inference 1 Introduction Recently in machine learning, relationallearning hasreceived a great deal of attention, for example, to findsocial roles in social network. While the stochasticblockmodel has been popular for such relational learn-ing in sociology, it is now widely used for various appli-cations including clustering proteins, discovering con-cepts, etc [9, 13].Infinite relational models (IRMs) [9, 13] are stochas-tic blockmodels exploiting the Dirichlet process [6, 2]sothatwedonotneedtodeterminethenumberofclus-ters a priori. Stochastic blockmodels of mixed mem-bership (SBMM) are also stochastic blockmodels thatmodel multiple observation of tables [1].In this paper, we propose a frequency-based infi-nite relational model (FIRM). The FIRM takes intoaccount the frequency of relation, which is statisti-cally informative but ignored by stochastic blockmod-els. Our model generalizes the IRM and can also ob-servemultipletablesastheSBMM.ToapplytheFIRMto a large dataset, we also derived a variational in-ference algorithm for the FIRM. Experimental resultsshow the FIRM gives better clustering results than theIRM on a dataset which has the frequency of relation.