Development of techniques for pre-processing and prediction of multi-label classification problems

Abstract

Introduction: Multi-label classification is a variant of traditional single-label classification, where an object is no longer classified by exclusively one label. Instead, this learning aims to assign one or more classes from a predefined set of classes to an object. Since multi-label learning is still in an early development stage compared to other classification techniques, some techniques currently available for other learning types have not been developed for this specific learning case.

Methods: After a survey of the existing literature, the following are some research challenges within this topic: data quality measures, reduction methods on multi-label datasets, outlier detection methods, pooling layers for multi-label data without a topological organization, methods to deal with multi-label classification problems with sparse features, and Explainable Artificial Intelligence techniques for multi-label neural classifiers.

Results: We propose: a) three measures of multi-label data quality, b) six methods for reducing multi-label datasets, c) a method that measures an object's anomaly degree in a multi-label dataset, d) a deep neural architecture using bidirectional association-based pooling layers, e) a neural system to solve multi-label classification problems described by tabular data that might involve sparse features, and f) an adaptation to the multi-label scenario of a classical post-hoc interpretability technique on neural networks. Conclusions, the proposed methods provide the scientific community with novel multi-label classification techniques, making possible a more efficient and effective knowledge discovery process on multi-label data.