Multilabel Rule Learning

点击翻译按钮获取中文摘要

Inductive rule learning is a very traditional, well-established resaerch area in machine learning. Rule learning algorithms are typically employed when one is not only interested in accurate predictions but also requires an interpretable theory that can be understood, analyzed, and qualitatively evaluated by domain experts. Ideally, by revealing the patterns and regularities that are implicitly captured in the data, a rule-based theory yields new insights in the application domain. On the other hand, in many machine learning tasks, predictions are sought for multiple target variables simultaneously, a problem known as multi-target prediction. Multilabel classification, where all output variables are binary, is an important special case. State-of-the-art methods in this area are able to improve performance by taking dependencies between these output variables into account. However, only little work has been done in learning explicit representations of such dependencies, which is a worth-while data mining task in itself.We are convinced that a rule-based view of this problem will greatly enhance our understanding and lead to better practical solutions. The main goal of this project is thus to connect research in multilabel classification and inductive rule learning, and to develop scalable rule learning algorithms for multilabel classification. Working in the intersection of two research areas in machine learning, we will make contributions to bothfields. At a very high level, the objectives of this project are thus (i) to develop a unified framework for representing different types of label dependencies, and analyze its expressive power for multilabel classification problems, (ii) to face the algorithmic challenges of learning multilabel rule sets from data, and (iii) to evaluate the predictive and descriptive performance of such rules in comparison to state-of-the-art systems.