Reasoning with unknown, not-applicable and irrelevant meta-values in concept learning and pattern discovery

摘要

This paper describes methods for reasoning with unknown, irrelevant, and not-applicable meta-values when learning concept descriptions from examples or discovering patterns in data. These types of meta-values represent different reasons for which regular values are not available, thus require different treatment in both rule learning and rule testing. The presented methods are handled internally, within the learning and testing algorithms, and not in preprocessing as those widely described in the literature. They have been implemented in the AQ21 multitask learning and knowledge discovery program, and experimentally tested on three real world and one designed datasets.