A study on partition quality of Fuzzy Co-clustering with exclusive item memberships
Katsuhiro Honda, Takaya Nakano, Seiki Ubukata, Akira Notsu · 2015
Bag-of-Words data analysis is a fundamental issue in web data mining for Big Data utilization, and Co-clustering is often applied to cooccurrence information analysis in such problems of document-keyword association research. In probabilistic partition models such as Multinomial Mixtures and Fuzzy c-Means-type ones, different partition constraints are forced to rows (objects) and columns (items), and then item memberships may not be useful in revealing item partitions. A possible approach in clarifying the interpretability of item partitions is additional penalization for exclusive item memberships, which was shown to emphasize cluster-wise representative items in document analysis. In this paper, the utility of the penalization approach is further studied through comparisons of partition qualities with several benchmark data sets. Several experimental results show that the additional penalty may sometime contribute to slightly improving the partition quality in addition to improvement of interpretability of co-cluster partitions.