Rapp, Michael (2022)
Multi-label Rule Learning.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00022099
Dissertation, Erstveröffentlichung, Verlagsversion

Kurzbeschreibung (Abstract)

Research on multi-label classification is concerned with developing and evaluating algorithms that learn a predictive model for the automatic assignment of data points to a subset of predefined class labels. This is in contrast to traditional classification settings, where individual data points cannot be assigned to more than a single class. As many practical use cases demand a flexible categorization of data, where classes must not necessarily be mutually exclusive, multi-label classification has become an established topic of machine learning research. Nowadays, it is used for the assignment of keywords to text documents, the annotation of multimedia files, such as images, videos, or audio recordings, as well as for diverse applications in biology, chemistry, social network analysis, or marketing. During the past decade, increasing interest in the topic has resulted in a wide variety of different multi-label classification methods. Following the principles of supervised learning, they derive a model from labeled training data, which can afterward be used to obtain predictions for yet unseen data. Besides complex statistical methods, such as artificial neural networks, symbolic learning approaches have not only been shown to provide state-of-the-art performance in many applications but are also a common choice in safety-critical domains that demand human-interpretable and verifiable machine learning models. In particular, rule learning algorithms have a long history of active research in the scientific community. They are often argued to meet the requirements of interpretable machine learning due to the human-legible representation of learned knowledge in terms of logical statements. This work presents a modular framework for implementing multi-label rule learning methods. It does not only provide a unified view of existing rule-based approaches to multi-label classification, but also facilitates the development of new learning algorithms. Two novel instantiations of the framework are investigated to demonstrate its flexibility. Whereas the first one relies on traditional rule learning techniques and focuses on interpretability, the second one is based on a generalization of the gradient boosting framework and focuses on predictive performance rather than the simplicity of models. Motivated by the increasing demand for highly scalable learning algorithms that are capable of processing large amounts of training data, this work also includes an extensive discussion of algorithmic optimizations and approximation techniques for the efficient induction of rules. As the novel multi-label classification methods that are presented in this work can be viewed as instantiations of the same framework, they can both benefit from most of these principles. Their effectiveness and efficiency are compared to existing baselines experimentally.

Frage zum Eintrag

Redaktionelle Details anzeigen