| Presentation | 2010-03-16 Feature Transformation Reflecting User's Relevance Takahiro TAKAMIYA, Toshikazu WADA, Shunji MAEDA, Hisae SHIBUYA, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | It is often pointed out by researchers working on similar image search that human and computer similarity measures are quite different. For solving this problem, metric learning method has been investigated based on relevance or preference scores given by users. This method is useful for some specific problems but cannot be combined with general classification problems. On the other hand, feature transformation methods, such as quantification method 4 and principal coordinate analysis that locates similar patterns closer and dissimilar patterns further. Since these methods use mutual similarity matrix for feature transformations, 1) they cannot transform input features without similarity scores to reference features, 2) an input feature transform causes feature transformations of all learnt data. Because of these limitations, these methods are mainly applied to data visualization tasks. In this report, we propose a feature transformation method free from these limitations by combining Relevance Component Analysis and quantification method 4. In the experiment, we applied SVM for classifying "Pen-Based Recognition of Handwritten Digits" in UCI repository and confirmed that our feature transformation improves the classification rate about 7%. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Feature transformation / Quantification method 4 / Relevance component analysis / k-means clustering |
| Paper # | PRMU2009-305,HIP2009-190 |
| Date of Issue |
| Conference Information | |
| Committee | HIP |
|---|---|
| Conference Date | 2010/3/8(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | |
| Vice Chair | |
| Secretary | |
| Assistant | |
| Paper Information | |
| Registration To | Human Information Processing (HIP) |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Feature Transformation Reflecting User's Relevance |
| Sub Title (in English) | |
| Keyword(1) | Feature transformation |
| Keyword(2) | Quantification method 4 |
| Keyword(3) | Relevance component analysis |
| Keyword(4) | k-means clustering |
| 1st Author's Name | Takahiro TAKAMIYA |
| 1st Author's Affiliation | Faculty of Systems Engineering, Wakayaam University() |
| 2nd Author's Name | Toshikazu WADA |
| 2nd Author's Affiliation | Production Engineering Research Laboratory, Hitachi,Ltd. |
| 3rd Author's Name | Shunji MAEDA |
| 3rd Author's Affiliation | Production Engineering Research Laboratory, Hitachi,Ltd. |
| 4th Author's Name | Hisae SHIBUYA |
| 4th Author's Affiliation | Production Engineering Research Laboratory, Hitachi,Ltd. |
| Date | 2010-03-16 |
| Paper # | PRMU2009-305,HIP2009-190 |
| Volume (vol) | vol.109 |
| Number (no) | 471 |
| Page | pp.pp.- |
| #Pages | 6 |
| Date of Issue | |