Estimating the factor structure in multiple matrices

2019 Fiscal Year Final Research Report

• Project/Area Number: 16H02868
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Research Field: Intelligent informatics
• Research Institution: Kyoto University
• Principal Investigator: Mamitsuka Hiroshi 京都大学, 化学研究所, 教授 (00346107)
• Project Period (FY): 2016-04-01 – 2019-03-31
• Keywords: 機械学習

Outline of Final Research Achievements

The objective of research is to present a generalized framework for the input of multiple matrices sharing dimensions and efficient solutions under this framework. We show two example results among our various results obtained during our research period:
1. For the input of a tensor and a matrix which share one dimension, we define a new norm and propose an efficient learning algorithm to estimate the norm. We analyze the property of the norm and empirically show the performance advantage of our norm and algorithm using both synthetic and real-world datasets. The results were summarized into a publication appeared in Neural Computation and also a paper appeared in NeurIPS, one of the top machine learning conferences.
2. We develop an efficient, scalable probabilistic-model based approach for the input of multiple matrices sharing dimensions. This result was published as a paper appeared in AAAI, one of the top artificial intelligence conferences.

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

次元を共有する複数の行列は、例えば、EC (electronic commerce) サイトの顧客（ユーザ）と商品（アイテム）のデータや、患者に対する薬の投与データで見受けられる。一般的な設定である。これらのデータに対する効率的・スケーラブルな機械学習手法は、将来的に多くの分野で利用される可能性があり、意義が大きいと考えられる。