Iku Ohama

An Extension of the Infinite Relational Model Incorporating Interaction between Objects

The Infinite Relational Model (IRM) introduced by Kemp et al. (Proc. AAAI2006) is one of the well-known probabilistic generative models for the co-clustering of relational data. The IRM describes the relationship among objects based on a stochastic block structure with infinitely many clusters. Although the IRM is flexible enough to learn a hidden structure with an unknown number of clusters, it sometimes fails to detect the structure if there is a large amount of noise or outliers. To overcome this problem, in this paper we propose an extension of the IRM by introducing a subset mechanism that selects a part of the data according to the interaction among objects. We also present posterior probabilities for running collapsed Gibbs sampling to learn the model from the given data. Finally, we ran experiments on synthetic and real-world datasets, and we showed that the proposed model is superior to the IRM in an environment with noise.

Discovering Relevance-Dependent Bicluster Structure from Relational Data

In this paper, we propose a statistical model for relevance-dependent biclustering to analyze relational data. The proposed model factorizes relational data into bicluster structure with two features: (1) each object in a cluster has a relevance value, which indicates how strongly the object relates to the cluster and (2) all clusters are related to at least one dense block. These features simplify the task of understanding the meaning of each cluster because only a few highly relevant objects need to be inspected. We introduced the RelevanceDependent Bernoulli Distribution (R-BD) as a prior for relevance-dependent binary matrices and proposed the novel Relevance-Dependent Infinite Biclustering (R-IB) model, which automatically estimates the number of clusters. Posterior inference can be performed efficiently using a collapsed Gibbs sampler because the parameters of the R-IB model can be fully marginalized out. Experimental results show that the R-IB extracts more essential bicluster structure with better computational efficiency than conventional models. We further observed that the biclustering results obtained by RIB facilitate interpretation of the meaning of each cluster.