Graph Mining for Automatic Classification of Logical Proofs
Karel Vaculík, Luboš Popeĺınský · 2014
We introduce graph mining for evaluation of logical proofs constructed by undergraduate students in the introductory course of logic. We start with description of the source data and their transformation into GraphML. As particular tasks may differ---students solve different tasks---we introduce a method for unification of resolution steps that enables to generate generalized frequent subgraphs. We then introduce a new system for graph mining that uses generalized frequent patterns as new attributes. We show that both overall accuracy and precision for incorrect resolution proofs overcome 97\%. We also discuss a use of emergent patterns and three-class classification (correct/incorrect/unrecognised).