Although knowledge discovery and data mining techniques have been successfully employed to various real-world applications, the classification of an imbalanced dataset remains challenging. This challenge has attracted attention from both academia and industry. Classification analysis is one of core research topics in pattern recognition field. According to the distribution of samples, algorithms like artificial network (ANN) and support vector machine (SVM) have been proposed to perform binary classification. But these traditional classification algorithms hardly work well for imbalanced dataset. This talk illustrates recent results in imbalanced data classification.

I Education
Dr. Junzo Watada received his B.Sc. and M.Sc. degrees in electrical engineering from Osaka City University, Japan, and Dr. of Eng. degree through the research on fuzzy multivariate analysis from Osaka Prefecture University, Japan.

II Profession
He is a Professor of Knowledge Engineering, Soft Computing and Management Engineering at the Graduate School of Information, Production & Systems, Waseda University since 2003, after a professor of Human Informatics and Knowledge Engineering, at the School of Industrial Engineering, the Osaka Institute of Technology, Japan. Also Dr. Watada gave as a fucclty lectures on Management Information Systems for 8 years at Faculty of Business Administration, Ryukoku University at Kyoto. Before moving to Academia, he was with Fujitsu Co. Ltd. as a senior systems engineer for 7 years.

III Research Interests
His research interests includes business decision making and management of technology and engineering as well as fuzzy system methodologies, automata theory, text and web mining, decision support systems and experts systems, DNA computing, data analysis, etc.. Recently he works actively on meta heuristics, and human tracking, as well as financial engineering He published more than 600 academic papers of journals and international proceedings.

IV Academic Contribution
He was the President of Bio-Medical Fuzzy Systems Association (2001-2003). He was the Vice President of Japan Society for Fuzzy Theory and Systems for two years (1993-1995) and was a board committee of Japan Society for Fuzzy Theory and Systems, and also serve as an advisory board member for several international and domestic societies and also an editorial board member for international and domestic journals including ICIC express letters.

V Awards
Dr. Watada received
1) The Contribution Award at ISIS2002 in Korea on August 25, 2001,
2) Henri Coanda Medal in Romania on July 17, 2002,
3) Excellent Presentation Award of SCIS2002 at Tsukuba on October 21-25, 2002,
4) Board of Certification in Professional Ergonomics as Certified Professional Ergonomist by Japan Ergonomics Society on August 3, 2003,
5) Contribution Award, Biomedical Fuzzy Systems Association, November, 2004,
6) Fellow, Society of Japan Intelligent Informatics and Fuzzy Systems, January 10, 2005, and
7) The Contribution Award, International Anniversary Symposium “Grigore C. Moisil” SASM2005, May 1-3, 2005
8) The Contribution Award, Japan Society of Fuzzy Theory and Intellectual Informatics, September 10, 2005,
9) The Contribution Award to developing fuzzy systems, on behalf of Professor L.A. Zadeh, BISC, Special Event 40 years of Fuzzy Systems, November 3, 2005.
10) Several best paper awards with his students

Abstract - Knowledge engineering is an important research field for intelligent systems. It includes any kind of data or knowledge preprocessing, learning, mining, and integration. In this speech, I will introduce how evolutionary computation can help improve the performance and quality of knowledge engineering. Three parts will be covered. In the first part, I will describe suitable integration techniques for various kinds of mined knowledge based on evolutionary computation. Some integration algorithms respectively derived from the Michigan approach and the Pittsburgh approach will be explained. The Michigan approach encodes each rule as an individual; on the contrary, the Pittsburgh approach encodes a rule set as an individual. Both of them have their own advantages and disadvantages. In the second part, I will introduce several GA-based fuzzy data-mining methods for automatically extracting membership functions for fuzzy association rules. All the genetic-fuzzy mining methods first use evolutional computation to find membership functions suitable for mining problems and then use the final best set of membership functions to mine fuzzy association rules. Through appropriately designed fitness functions, these approaches can avoid the formation of bad kinds of membership functions and can provide important mining results to users. In addition, feature selection is an important pre-processing step in mining and learning. A good set of features can not only improve the accuracy of classification, but also reduce the time to derive rules. Thus in the last part, I will state some GA-based clustering methods for attribute clustering and feature selection. The proposed approaches for attribute clustering can also easily handle the problem of missing values in classification.

Global and local methods in face representation and recognition

Face recognition has received more and more attention. The automatic face recognition technology can be applied to personal authentication, access control, check on work attendance, etc. However, this technology is greatly influenced by varying poses, illumination and facial expression, as well as aging variation. This requires that robust face representation and recognition methods be proposed. The majority of face representation and recognition methods can be grouped into two kinds, global methods and local methods. Global methods exploit the global structure and information inside the face images to recognize faces. Typical examples of global methods include principal component analysis (PCA), linear discriminant analysis (LDA), and so on.  Most of global methods are in the scope of statistical methods and are trained by only training samples, which are only a few sampling results of real face images. Consequently, the statistical methods are good at representing the variation of face images that have been observed, but is not competent in representing “unseen” variation of the face images. The test sample can be somewhat viewed as “unseen” variation of the face images, especially in the complex conditions. Thus, it seems that we should not desire that the global methods bring us a surprise in face recognition performance. There are two kinds of local methods. The first kind of local methods try the best to use the local structure and information of training samples. These methods are also in the scope of statistical methods. Typical examples include Graph embedding methods, locality preserving projection (LPP), manifold learning. The second kind of local methods attempts to exploit “local” training samples to provide representation for the test sample. Typical examples include the sparse representation method (SRM). SRM usually exhibits a good face recognition performance, but it appears that people cannot provide an explanation for the performance from the viewpoint of applications. In order o make more progresses in face recognition, we have to explore and do more. At the moment, to develop computationally efficient and theoretically explainable SRMs and to construct a bridge for SRM and global methods are two things that are worth doing.