White-box data mining algorithms

GRAND Seminar March 26, 3PM, Monday, 2012, ENGR 3507

Boris Delibasic
Associate professor
University of Belgrade in Serbia

Host:

Alex Brodsky

Abstract:

Choosing the right algorithm for data at hand was always a major problem in data mining. We propose a new architecture for decision-support systems for data mining, with the ability of generic algorithm design to help users choose the right algorithm. Opposite to the prevalent black-box approach of using algorithms in data mining were users have the ability to define inputs, setup parameters and read outputs, we propose using reusable component (RC) based algorithms. The RC-based algorithms are assembled from reusable components, which are standalone algorithm units which were originally found in black-box algorithms and their partial improvements. RC based algorithms have been proven to better adapt to data than black-box algorithms that, due to “hard” bindings of algorithm parts, are disabled to achieve best results on some datasets. On the other hand, the RC-based approach of algorithm design produces a galore of algorithms making it thus harder to search through the algorithm space. We show how this problem can be solved using meta-heuristics for searching through the algorithm space. We also propose further research directions that will enable to connect the proposed approach with meta-learning. We believe that users will be better supported in the future for choosing an adequate algorithm for the problem at hand, because the decision support system will be enabled to perform an intelligent search through the algorithm space that is based on dataset properties, algorithm performance results, empirical rules gained from meta-learning and theoretical support.

Short Bio:

Boris Delibasic is an associate professor at the University of Belgrade in Serbia. His main research interests are data mining, decision support systems, business intelligence, and decision theory. Dr. Delibasic is also an adjunct lecturer at the University of Jena in Germany. He has already published several research articles in top-ranked international journals. A project he is currently engaged with is dealing with design of white-box algorithms for data mining (www.whibo.fon.bg.ac.rs). In 2011, Prof. Delibasic received a prestigious Fulbright fellowship to work as a visiting scholar at Zoran Obradovic’s Center for data analytics and biomedical informatics at Temple University in Philadelphia, PA. His current research objectives are to design spatio-temporal algorithms for analysis of ski injuries and to discover ski injury patterns that could be used for injury prevention. Algorithms developed for ski injury analysis, are planned in a later stage to be extended, to analyze large scale data on road traffic accidents. Dr. Delibasic is also ski patroller on Serbian mountains during the winter season.