Learning Decision Rules or Learning Decision Models?

Abstract

In this position paper, we discuss the reasons for the lack of success of rule learning, as witnessed by the quasi-absence of commercial applications, and what can be done to revive the domain and possibly to kick-start the kind of explosive development that statistical Machine Learning and Neural Networks have experienced over the past 15 years. The root cause of the problem is well-known, and it is not the rule learning algorithms themselves: if the representation language to which a rule learning algorithm has access – often only the representation model of the input data – is not appropriate to represent the decision rules, the algorithm has no way to generate a decision ruleset that generalizes well. Feature generation and other methods that have been proposed to augment the data representation language are useful, but we argue that the focus should be on discovering the conceptual model that underly the decision. We claim that this amounts to discovering the structure of decisions, that is, to learn decision models. We outline some potentially fruitful research directions, and how this topic is central to neuro-symbolic learning.