AutoML for Multi-Label Classification: Overview and Empirical Evaluation

Automated machine learning (AutoML) supports the algorithmic construction and data-specific customization of pipelines, including selection, combination, parametrization algorithms as main constituents. Generally speaking, AutoML approaches comprise two major components: a search space model an optimizer for traversing space. Recent have shown impressive results in realm supervised learning, most notably (single-label) classification (SLC). Moreover, first attempts at extending these towards multi-label (MLC) been made. While candidate pipelines is already huge SLC, complexity raised to even higher power MLC. One may wonder, therefore, whether what extent optimizers established SLC can scale this increased complexity, how they compare each other. This paper makes following contributions: First, we survey existing Second, augment with not previously tried Third, propose benchmarking framework that fair systematic comparison. Fourth, conduct extensive experimental study, evaluating methods on suite MLC problems. We find grammar-based best-first favorably other optimizers.