Multi-label learning on low label density sets with few examples

Multi-label learning has experienced an immense growth in the last years due to the multiple real-life applications to which it is applicable, such as the classification of protein functions, or musical genres, among others. This has led to the proposal of categories for multi-label classification (MLC) problems that seek to establish guidelines for the different configurations, given either by the quality or quantity of the labels, the number of examples for training, etc. Such is the case for the class of problems known as "Challenging MLC", those in which the universe of labels incorporates obstacles either in terms of quality (erroneously assigned labels, unseen labels, etc.) or quantity (thousands or millions of labels). Different methods have been developed to address these cases, and yet few efforts have been directed towards the case where, despite having a large label universe, the number of examples is small (of the same order as the labels), thus posing amore complex scenario. In this paper, we examine one important real-world problem case - the labeling of Geometric surface patterns, appearing on pottery objects from the Classical era. As we will show, existing methods from the state of the art can provide baseline performance, but cannot yet comprehensively address this and similar application problems. We present and encompassing experimental comparison of state of the art methods, detailing advantages and problems. We contribute a processing pipeline that allows us to achieve effective classifications. Our work addresses the importance case when the universe of labels admits a feasible simplification through natural language processing (NLP) techniques and augmentation of visual training data. Based on an in-depth analysis of results, we propose practical guidelines on how to face similar problems, regarding both the selection of techniques and the analysis of results. We also identify pressing issues for current research to make multi-labeling more widely applicable and functional.