Relevance-dependent Biclustering Models for Relational Data Analysis
郁 大濵 · Hokkaido University Collection of Scholarly and Academic Papers (Hokkaido University) · 2019
Relational data encoding pairwise relationships between objects appears in many fields.For example point-of-sale (POS) data of an e-commerce (EC) site contain relational data between customers and items, and follower lists in social networking services (SNS) such as Twitter is relational data among users.Recently, with the rapid advancements in internet technologies, a large amount of relational data has been accumulated in many business fields.Therefore, extracting insights by analysing relational data becomes an important challenge for many business persons to refine their business activities.Biclustering is one of the most popular techniques to extract useful insights from relational data.Biclustering abstracts the given data matrix into a low-dimensional block structure by simultaneously clustering both the row and column objects.For extracting robust bicluster structure from noisy real-world relational data, there have been studied many statistical models for biclustering.Among these models, the Infinite Relational Model (IRM) proposed by Kemp et al. is one of the most fundamental biclustering models.The IRM abstracts given relational data into a block structure, in which each block has its own link probability.The IRM can automatically estimates the optimal number of clusters.Furthermore, posterior inference for the IRM can be performed efficiently using collapsed Gibbs sampler.The IRM and its extended models commonly assume that each block of the bicluster structure has an uniform density.However, this assumption is not acceptable in many i I would like to thank everyone in Information Knowledge Network laboratory, especially Ms. Yu Manabe.Her supports significantly facilitated my research activities.