2005 Fiscal Year Final Research Report Summary
Development of Classification Systems for Large-Scale Pattern Recognition Problems and Its Application
| Project/Area Number |
14380151
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (B)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | HOKKAIDO UNIVERSITY |
|---|
Principal Investigator |
KUDO Mineichi Hokkaido Univ., Grad.School of Info.Sci.and Tech., Prof., 大学院・情報科学研究科, 教授 (60205101)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
MURAI Tetsyua Hokkaido Univ., Grad.School of Info.Sci.and Tech., Assi.Prof., 大学院・情報科学研究科, 助教授 (90201805)
IMAI Hideyuki Hokkaido Univ., Grad.School of Info.Sci.and Tech., Assi.Prof., 大学院・情報科学研究科, 助教授 (10213216)
TOYAMA Jun Hokkaido Univ., Grad.School of Info.Sci.and Tech., Instructer, 大学院・情報科学研究科, 助手 (60197960)
TENMOTO Hiroshi Kushiro National College of Tech., Dep.of Info.Eng., Assi.Prof., 情報工学科, 助教授 (80321371)
HAYASHI Hiroki Kushiro National College of Tech., Dep.of Info.Eng., Assi.Prof., 情報工学科, 助教授 (60342440)
|
|---|
| Project Period (FY) |
2002 – 2005
|
| Keywords | pattern recognition / large-scale problem / feature selection / generalization / class number / subclass / super-class / visualization |
|---|
| Research Abstract |
In this study, we classified the scaling problems of pattern recognition tasks into three of the following. The results are shown below, respectively.
1.Scaling problem as for data number : We have changed the problem setting from the problem to achieve as high (predictive) performance as possible to the problem to attain as least computational time as possible keeping no large damage in performance. Especially, 1)a fast k-nearest neighbor was invented, and 2)a one-pass algorithm for prototype acquisition was discussed, though they are still under progress.
2.Scaling problem as for dimensionality : We concentrated on "feature selection" which removes low-informative features but keeps the classification performance. Solving this problem usually require a combinatorial examination, so many sub-optimal approaches have been proposed so far. However, they are not satisfactory in time for large-scale problems in the number of original features. In this study, we focused on "classifier-indepen
… More
dent feature selection" in which only garbage (completely no informative for classification) features are removed. Usual algorithms do "classifier-dependent feature selection." We showed that the former can be made faster than the latter, and a two-stage selection scheme (CIFS+CSFS) is efficient for feature selection.
3.Scaling problem as for cass number : As seen in Kanji-Character recognition, we sometimes have to deal with many categories. The difficulty of classification task increases as the class number increases. We proposed a way to analyze the relationship among classes and grouped some hard-to-classify classes into one "super class." In addition, we investigated a visualization method using a graph representation and a tree representation for making clear the relationship. In the decision tree, we confirmed that the total performance can be increased with feature selection. Last, we suggested that such super-classes and subclasses analysis is useful to capture the individual classification problems. Less
|
