University of Konstanz
Graduiertenkolleg / PhD Program
Computer and Information Science

Graduation Talks

title

Learning Hierarchical Fuzzy Rule Systems

speaker

Thomas Gabriel, University Konstanz
Konstanz, Germany

date & place

Wednesday, 09.07.2008, 16:15 h
Room C252

abstract

In data mining applications, rule systems are widely applied to learn human readable and understandable models from huge amounts of data. However, learning classical rule systems from real-world data often generates too many rules for human interpretation or rules which are too simplistic to be useful. In contrast, hierarchical rule systems with only a varying, small number of rules at each level can help to overcome this issue: Lower levels in the hierarchy focus on areas of the feature space where only weak evidence for a rule is found in the data. Rules further up at higher levels of the hierarchy describe increasingly general and strongly supported aspects of the data. In this thesis, we developed a base fuzzy rule learning algorithm to generate a hierarchical rule system. This approach leads to a well-defined arrangement of hierarchy levels allowing visual interactions and exploration across the different rule models. The rule hierarchy is evaluated on a number of benchmark datasets from the UCI machine learning repository: A better classification accuracy can be achieved where at the same time the number of rules decreases significantly.
The underlying fuzzy rule algorithm along with an intelligent filter strategy allows the generation of such a hierarchy. After the first rule model is trained, rules with low relevance are determined and are arranged in one hierarchy level of abstraction. This rule level is used in the next step as a filter for the training data before the next training phase starts. Training examples that resulted in the creation of small, less important rules are therefore excluded from the training phase of the next layer, resulting in a more general rule system ignoring the small withheld details in the training data. The generated rule hierarchy provides more accurate results compared to a non-hierarchical rule learner, and at the same time uses less rules which also allows visual exploration across different levels of the hierarchy.
Two visualization techniques are extended to display rule hierarchies and data points in one picture. It is demonstrated how a connected set of Parallel Coordinates can be used to visually explore this hierarchy of rule systems and allows interactive brushing of the underlying model hierarchy. A second approach for visualizing high-dimensional points known as Multi-Dimensional Scaling, is applied to map rule centers, corresponding coverages, and training data on one coherent 2D plot.