This dissertation investigates rule learning as a statistical classification problem. Rule learning is known for inducing interpretable and comprehensible classifiers, whereas statistical machine learning has often focused on predictive accuracy. The goal of this work is to combine the approaches to obtain highly predictive, yet interpretable classifiers. To this end, we investigate two types of rule sets. For DNF formulae, we propose to use stochastic local search with structural risk minimization or ensemble-based capacity control. For weighted rule sets, we devise the novel convex optimization criterion “Margin Minus Variance” and derive a new concentration inequality for capacity control. Finally, for multi-relational learning, we propose a framework to assess and compare learning systems according to the flow of information and present a novel rule generation criterion that aims at highly diverse rule sets.
«
This dissertation investigates rule learning as a statistical classification problem. Rule learning is known for inducing interpretable and comprehensible classifiers, whereas statistical machine learning has often focused on predictive accuracy. The goal of this work is to combine the approaches to obtain highly predictive, yet interpretable classifiers. To this end, we investigate two types of rule sets. For DNF formulae, we propose to use stochastic local search with structural risk minimizat...
»