- [Home]
- [Research achievement]
- [Research achievement detail]
Title | Hierarchical Multi-label Classification on Statistical Decision Theory (in Japanese) |
---|---|
Authors | 山本粋士 、須子統太 、松嶋敏泰 |
Released Year | 2012 |
Format | Conference |
Category | Knowledge information processing |
Jounal Name | |
Jounal Page | vol.112, no,454, pp.101-106 |
Published Year | 2013 |
Published Month | 2 |
Abstract (English) |
This paper considers multi-label classification on statistical decision theory. In Label Power Set format, multi-label classification is equivalent to multi-class classification. However, the number of classes increases exponentially as elements in label set grow in number. Hence in case of many labels, a prohibitive computational cost problem occurs. To avoid this problem, some studies have been done and one of them used hierarchical structure. On the other hand, optimal classification method based on bayes rule has been attracted much attention recently. We apply this optimal classification method based on bayes rule to multi-label classification problem. Moreover, assuming hierarchical structure on labels, we propose efficient classification algorithms which reduce computational cost to linear order on the number of elements in label set. Since optimal classification based on bayes rule differs calculation formula depending loss function, we present algorithms in case of O-1 loss and hamming loss, respectively. |
Note (English) |
1 |
Manuscript | |
Presentation |
Involved Papers
- A Consideration on Minimum Storage Regenerating Codes for Functions (in Japanese)
- A Note on the Inference Algorithm on the Factor Graph based on the Linear Programming (in Japanese)
- An Efficient Algorithm for the Bayes Codes using an Unlimited Depth Context Tree
- Linear Programming Decoding of Binary Linear Codes for Multiple-Access Channel (in Japanese)
- A Note on the Fixed-Length Source Coding Theorems for Sources with Side Information (in Japanese)