Improved Comprehensibility and Reliability of Explanations via Restricted Halfspace Discretization
MLDM '09 Proceedings of the 6th International Conference on Machine Learning and Data Mining in Pattern Recognition
| Hi-index | 0.00 |
The problem of finding compact logic explanations subsumes the separation problem of machine learning, where one desires a mathematical formula or, more generally, an algorithm that decides membership of the elements of two or more given data sets. We derive a solution to the problem where the explanations are represented as propositional logic clauses. These clauses act as separation rules uncovering hidden relationships in the underlying data. The method for finding the explanations consists of several steps including discretization, introduction of uncertainty, computation of importance functions, construction of explanation formulas, and validation. The first three steps rely on a new notion of alternate random process, which is essential for identification of salient features and relationships. In several tests, the process produced compact logic explanations that were easily understood and readily verified as correct or at least potentially correct by a field expert.