Intelligent Data Analysis and Graphical Models
Dr. Christian Borgelt
The main objectives of the Intelligent Data Analysis and Graphical Models Research Unit are the development of intelligent methods for analyzing data, as it is collected and stored nowadays pratically everywhere, and the application of graphical models (a special model for handling uncertainty that uses graph representations) for modeling, diagnosis, and planning.
Due to modern information technology, which produces ever more powerful computers every year, it is possible today to collect, transfer, combine, and store huge amounts of data at very low costs. Thus an ever-increasing number of companies and scientific and governmental institutions collects data in electronic form. However, even though any single bit of information can be retrieved and simple aggregations can be computed, general patterns, structures, and regularities often go undetected. In order to find these patterns and thus to exploit more of the information contained in the available data, we need intelligent tools that help us to transform data into useful knowledge. The gained insights may then be exploited to increase turnover and profit, to better the product quality, or to improve customer satisfaction. This research unit develops and implements a large variety of intelligent data analysis and data mining methods, ranging from decision and regression trees, Bayes classifiers, and artificial neural networks to clustering, frequent pattern mining, and trend detection.
Graphical models are a special tool for handling uncertainty in complex domains, which have grown more and more popular in the last 15 years. Earlier approaches tried to extend classical, logic-based methods by simple means to handle uncertainty, but failed, because they implicitely make assumptions about conditional independences that are rarely satisfied. Graphical models tackle this problem by explicitely representing all valid conditional independences in a graph, thus providing a sound and well-founded methodology for uncertainty handling. Their area of application is almost unlimited and includes knowledge representation, system modeling, medical and technical diagnosis, and planning support. Methods for learning graphical models from data aid a user in the task to develop an appropriate model for the domain under consideration.
