| The traditional clustering algorithms only analyze the data from a single viewand the clustering lacks semantic labels.Thus the clustering is not interpretable.Multi-view means clustering based on multiple views.Interpretability is that interpretable qualitatively and quantitatively clustering result is obtained by decision rules.Multi-views’ and interpretable clustering provides users with more choices.Thus users can use and trust clustering results with critique,improvement and exploration.In this thesis,we propose an Interpretable Clustering with Multi-views Generative Model(ICMG)to solve the problems of multi-view and interpretability.ICMG can generate multiple views,and obtain a number of effective and non-redundant clusterings based on the views.Finally,ICMG interprets clustering qualitatively and quantitativelythrough the semantic information of the views.The work done in this thesis is as follows.(1)We construct a Multi-view Bayesian Case Model(MBCM).MBCM is a generative model that introduces multiple views into the Bayesian Case Model,and the data generated by MBCM includes multiple views.(2)We construct a Multi-view Generative Model(MGM).MGM generates a number of effective and non-redundant views based on the principle of efficiency and non-redundant by using the combination of Bayesian Program Learning(BPL)and MBCM.MGM uses prototypes and subspaces to describe views.(3)We propose an Interpretable Clustering with Multi-view Generative model(ICMG).ICMG firstly uses MGM to obtain multiple views descripted prototypes and subspaces.Secondly,ICMG constructs the set of rules using the prototypes and subspaces and then ICMG clusters data based on the rules.Finally,ICMG explains clustering qualitatively and quantitatively by using the semantic information of prototypes and subspaces,then ICMG gets semantic class labels.The thesis uses many groups of data sets in the experiment.The experimental results show that ICMG can get multiple interpretable clustering results,and the clustering performance is superiorto traditional multi-view clustering algorithms.The experiment indicates users can easily understand the ICMG’s clustering results than the traditional multi-view clustering results. |
