Improvement of Nonnegative Matrix Factorization method using competitive learning

2016 Fiscal Year Final Research Report

• Project/Area Number: 26330259
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Intelligent informatics
• Research Institution: Hokkaido Information University
• Principal Investigator: Uchiyama Toshio 北海道情報大学, 経営情報学部, 教授 (80708644)
• Project Period (FY): 2014-04-01 – 2017-03-31
• Keywords: 競合学習 / 情報理論的クラスタリング / 非負値行列因子分解 / トピックモデル

Outline of Final Research Achievements

This study investigated novel methods to improve the accuracy of nonnegative matrix factorization (NMF) from both theoretical and experimental side. Theoretically, it has shown the equivalence between information-theoretic clustering (ITC) and NMF based on generalized KL divergence. Then, it proposed a novel initialization method for NMF using ITC and experimentally showed the effectiveness of the method compared to conventional methods. It also proposed another algorithm for NMF using competitive learning which selects a subset of vectors as winner and showed the effectiveness.

Free Research Field

データマイニング