Integrated Learning of Supervised and Unsupervised Learning : Psychological Experiments and Computational Model

Abstract

In this paper, we describe psychological experiments for integration of unsupervised learning and supervised learning. This integrated learning has an advantage over simple unsupervised learning since it outputs a class which is explained by a useful classifier. However, little attention has been paid to this kind of learning. In the integrated learning, an overly complex class gives little information and a bad classifier. It is important to consider complexity of a class and goodnes of a classifier and select attributes in clustering. Therefore, attribute selection criteria in clustering has a crucial effect on the results. We conduct psychological experiments with a commercial date set in order to investigate human beings' process of building the criteria and discover useful knowledge. These experiments show that the criteria which consider complexity or property of a class are as good as the best criterion which requires domain knowledge. Based on these experiments, we build a domain-independent computational model for integrated learning. This model was applied to an agricultural data set, and the results were appreciated by the experts. Although the obtained class is relatively complex, the classifier has an accuracy of 89%.