R. Kamimura (Japan)
: mutual information maximization, commonfeatures, distinctive features, cost, cost-sensitive
In this paper, we propose information-theoretic methods to decompose input patterns into basic features. By de composing input patterns, it is possible to interpret clearly the main characteristics of input patterns. Basic features are supposed to be composed of common and distinctive features. Distinctive features are expected to be extracted by information maximization. To detect common features, we introduce a cost in the framework of information max imization, which is called cost-sensitive information max imization. The cost is defined so as to take into account averaged input patterns for common feature detection. We applied our method to an artificial data problem and stu dents' questionnaire analysis. Experimental results show that common and distinctive features are clearly separated by our method. In addition, we could see that the main characteristics of data can clearly be interpreted.
Important Links:
Go Back
