Autonomous Clustering for Machine Learning

In this paper, starting from a collection of training examples, we show how to produce a very compact set of classification rules. The induction idea is a clustering principle based on Kohonen’s self-organizing algorithms. The function to optimize in the aggregation of examples to become rules is a classificatory quality measure called impurity level, which was previously employed in our system called FAN. The rule conditions obtained in this way are densely populated areas in the attribute space. The main goal of our system, in addition to its accuracy, is the high quality of explanations that it can provide attached to the classification decisions.