Learning Complexity-Bounded Rule-Based Classifiers by Combining Association Analysis and Genetic Algorithms
Yugen Yi, Eyke Hüllermeier · 2005
We propose a method for learning rule-based classifiers that can be seen as a compromise between a complete enumeration of the hypothesis space and heuristic strategies to search this space in a greedy manner. The method consists of two main parts: In a first step, a sufficiently large number of individual, high-quality classification rules is generated. In a second step, a classifier is assembled from the candidate rules thus obtained. This comes down to selecting a proper subset of these rules, a combinatorial problem that we shall approach by means of genetic algorithms. For the candidate generation step, we suggest using association rule mining. Apart from learning accurate classifiers, a main motivation of our method is the possibility to control the tradeoff between accuracy and transparency (complexity) of a model in a more explicit way.