Computational intelligence tools
for data understanding.


Web page in preparation for a
tutorial, to be presented on 29th June, 9 am -12 am, at:

Joint International Conference On Artificial Neural Networks and International Conference on Neural Information Processing Conferences (ICANN/ICONIP 2003), Istanbul, Turkey, 26-29.06, 2003

UMK - logo


Wlodzislaw Duch

wduch

Department of Informatics,
Nicolaus Copernicus University,
Torun, Poland


What this tutorial is about ?

This tutorial presents computational intelligence approach to data mining, stressing the need for understanding of the data structure. At each step computer programs will be used in real-time on real-world examples to illustrate various procedures involved. Following issues are covered:

  1. Relations between AI, CI and Data Mining.
  2. Forms of useful knowledge, including logical rules (crisp and fuzzy), decision trees, prototype-based rules and visualization techniques. The need for, and advantages of, various types of data analysis is explained.
  3. A short description of the philosophy of integration of algorithms used in our GhostMiner software is presented, including an outline of algorithms used in the software: neural, neruofuzzy, decision tree, similarity based, SVM, committees of models and MDS visualization.
  4. Exploratory data analysis and visualization of proximity of data vectors.
  5. Rule-based data analysis: crisp and fuzzy logical rules are extracted and a tradeoff between accuracy/simplicity is explained using logical rules generated by neurofuzzy and decision tree models.
  6. Evaluation of results, error functions, ROC curves, and optimization of logical rules extracted from data, exploring the tradeoff between rejection/error level.
  7. Calculation of classification probabilities from any black-box system is presented. Assuming Gaussian uncertainties of measurements and crisp logical rules leads to analytical formulas that allow to optimize large complex sets of logical rules using gradient procedures. In effect interpretation is easy and accuracy is high.
  8. Prototype rules, similarity-based models and generation of prototype rules using decision tree.
  9. Visualization of decision borders of classifiers is presented as an alternative method of data understanding, visualization of neural network decisions.
  10. Lessons from applications of this approach to a few real life problems are analyzed, and simple logical rules for many datasets provided.
  11. A real life example of data mining for logical rules that are used in expert system.

References:

Duch W, Setiono R, Zurada J, forthcoming
Duch W, Adamczak R, Grabczewski K, Methodology of extraction, optimization and application of crisp and fuzzy logical rules. IEEE Transactions on Neural Networks, 12 (2001) 277-306

Link to the PowerPoint presentation of the tutorial.


CV of the main presenter:

Wlodzislaw Duch is an editor of a number of professional journals, including IEEE Transactions on Neural Networks, Computer Physics Communications, heads the scientific committee of the Polish "Cognitive Science" journal, is a member of the European Neural Network Society (ENNS) executive committee, Board of Governors of the International Neural Network Society (INNS), and Technical Committee of the IEEE Neural Network Society (NNS). He is a professor of applied computational sciences and theoretical physics, since 1990 heading the Department of Informatics (formerly called a Department of Computer Methods) at Nicolaus Copernicus University, Torun, Poland. He runs a software company DuchSoft, and was an executive board president of the e-business Kopernik.pl company. He holds the title of a "professor of theoretical physics and applied computational sciences" (1997), has habilitation degree in many body physics (D.Sc. 1987), and Ph.D. in quantum chemistry (1980). In the last 10 years his scientific interest moved from computational methods of physics, chemistry, quantum computing, to artificial intelligence, neural networks, medical applications of computational intelligence and various aspects of cognitive science.

His current interests are in computational intelligence (CI) methods, especially methods that elucidate the structure of the data, visualization methods, medical and psychological diagnosis support, methods that are applicable to complex objects, bioinformatics, and modeling the brain functions at different levels. His ambitions are to create general CI theory based on similarity evaluation, create meta-learning schemes that automatically discover the best model for a given data, popularize a new form of logical rules based on similarity to prototypes, develop geometrical theories for modeling of mental events and relating such models to neurodynamics.

In 2003 W. Duch works at the Nanyang Technological University for his sabbatical year. He has held a number of academic positions at universities and scientific institutions all over the world. These include University of Southern California in Los Angeles and the University of Florida in Gainesville, USA, University of Alberta in Edmonton, Canada, Meiji University, Kyushu Institute of Technology and Rikkyo University in Japan, Louis Pasteur Universite in Strasbourg, France, Max-Planck-Institut für Astrophysik in Germany (every year between 1984-2001), King's College London in UK, to name only a few.

He worked as an expert for the European Union 5th and 6th Framework science programs, Polish Committee of Scientific Research and the Ministry of Education. He has published 4 books and over 250 scientific and popular articles in many journals. He has been awarded a number of grants by Polish state agencies, foreign committees as well as European Union institutions.

His full CV is here.