Multi-Layered Framework for Modeling Relationships between Biased Objects
Iku Ohama, Takuya Kida, Hiroki Arimura · 2015
Latent variable models for relational data enable us to extract a co-cluster structure underlying observed relational data. The Infinite Relational Model (IRM) is a well-known relational model for discovering co-cluster structures with an unknown number of clusters. The IRM and several related models commonly assume that link probability between two objects depends only on their cluster assignment. However, relational models based on this assumption often lead us to extract many non-informative and unexpected clusters. This is because the cluster structures underlying real-world relationships are often blurred by biases that are inherent to individual objects. To overcome this problem, we propose a multilayered framework that extracts a clear co-cluster structure in the presence of objects' biases. Then, we propose a new model which is a special instance of the proposed framework that incorporates the IRM. Furthermore, we reveal that some relational models can be regarded as special cases of the proposed model. We present an efficient Gibbs sampler for posterior inference. Experiments conducted using real-world datasets confirm that the proposed model successfully extracts clear and interpretable cluster structures from blurred relational data.