Inductive Hypothesis Validation and Bias Selection in Unsupervised Learning

This paper approaches the importance of bias selection in the context of validating Knowledge Bases (KB) obtained by inductive learning systems. We propose a framework for automatic validation of induced KBs based on the capability of shifting the bias in the inductive learning system. We claim that this framework is useful not only when the system has to validate its own results, but also when human experts are available to perform the validation process. Experiments are made using accuracy in attribute prediction as performance goal. The unsupervised inductive learning system ISAAC [TC96] is coupled with the wrapper method [JKP94] to search for the best bias. The results support the proposed ideas and suggest some future work that seems interesting from both KB validation and Machine Learning points of view.