Presentation | 2018-03-05 Transformed Multiple Matrix Factorization: Towards Utilizing Heterogeneous Auxiliary Information Taira Tsuchiya, Tomoharu Iwata, Tetsuji Ogawa, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | Matrix factorization is widely used for a variety of fields, such as computer vision, document analysis, signal processing, and collaborative filtering. Matrix factorization acquires latent representations for rows and columns so that their inner products can represent elements of a target matrix. Auxiliary matrices with different properties are often available. Multiple matrix factorization exploits these auxiliary matrices to make matrix factorization more robust against the sparseness of the target matrix. Multiple matrix factorization, however, fails to utilize auxiliary information when these matrices have heterogeneous relationships against the target matrix because it learns the identical latent representations for multiple matrices, resulting in modeling linear relationships between matrices. In the present paper, transformed multiple matrix factorization is proposed to tackle this problem. The proposed method assumes that common latent representations, which are newly introduced, are shared across different matrices, but they are transformed by different non-linear functions to obtain latent representations for each matrix. The proposed method can therefore handle heterogeneous relationships among different matrices. Experiments conducted on three real-world datasets demonstrated that the proposed method outperformed vanilla matrix factorization and multiple matrix factorization. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | matrix factorizatoin / multi-task learning / transfer learning / neural networks |
Paper # | IBISML2017-96 |
Date of Issue | 2018-02-26 (IBISML) |
Conference Information | |
Committee | IBISML |
---|---|
Conference Date | 2018/3/5(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Nishijin Plaza, Kyushu University |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Statisitical Mathematics, Machine Learning, Data Mining, etc. |
Chair | Kenji Fukumizu(ISM) |
Vice Chair | Masashi Sugiyama(Univ. of Tokyo) / Hisashi Kashima(Kyoto Univ.) |
Secretary | Masashi Sugiyama(Nagoya Inst. of Tech.) / Hisashi Kashima(Univ. of Tokyo) |
Assistant | Tomoharu Iwata(NTT) / Toshihiro Kamishima(AIST) |
Paper Information | |
Registration To | Technical Committee on Infomation-Based Induction Sciences and Machine Learning |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Transformed Multiple Matrix Factorization: Towards Utilizing Heterogeneous Auxiliary Information |
Sub Title (in English) | |
Keyword(1) | matrix factorizatoin |
Keyword(2) | multi-task learning |
Keyword(3) | transfer learning |
Keyword(4) | neural networks |
1st Author's Name | Taira Tsuchiya |
1st Author's Affiliation | Waseda University(Waseda Univ.) |
2nd Author's Name | Tomoharu Iwata |
2nd Author's Affiliation | Nippon Telegraph and Telephone Corporation(NTT) |
3rd Author's Name | Tetsuji Ogawa |
3rd Author's Affiliation | Waseda University(Waseda Univ.) |
Date | 2018-03-05 |
Paper # | IBISML2017-96 |
Volume (vol) | vol.117 |
Number (no) | IBISML-475 |
Page | pp.pp.41-48(IBISML), |
#Pages | 8 |
Date of Issue | 2018-02-26 (IBISML) |