HyCASTLE: A Hybrid ClAssification System based on Typicality, Labels and Entropy

• Published in: Knowledge-based systems 2022-05, Vol.244, p.108566, Article 108566
• Main Authors: Delli Veneri, Michele, Cavuoti, Stefano, Abbruzzese, Roberto, Brescia, Massimo, Sperlì, Giancarlo, Moscato, Vincenzo, Longo, Giuseppe
• Format: Article
• Language: English
• Subjects: Classification; Clustering; Hybrid models; Hybrid systems; Iterative methods; Labels; Machine learning; Multi-class classification; Performance evaluation; Supervised classification; Topology
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2022.108566

Traditional supervised classification models aim to approximate the functional mapping between instance attributes and their class labels. These models, however, do not consider the interdependence between instances and global characteristics of data and thus often they lead to poor classification results. In this work, we present a novel hybrid classification model – named HyCASTLE – designed to solve the main shortcomings of hybrid models that employ topological information through clustering in order to improve classifiers performances: they make hypotheses on the underlying data distribution and do not consider the effect of noise. HyCASTLE utilises a non-parametric estimator to capture the underlying data distribution and creates entirely data-driven shape-free clusters. HyCASTLE then refines this cluster configuration using both data topology and available labels through an iterative cluster aggregation and separation process. We evaluated HyCASTLE performance on 37 datasets and compare it with both traditional and hybrid classification models. Our results show that HyCASTLE has comparable or better performance than the other models and results to be more resilient to class noise. •HyCASTLE produces shape-free clusters after outlier removal from the training set.•It considers both the data topological properties and labels to form clusters.•Dynamic topological constraints are used to minimise clusters entropy.•All cluster information is used to minimise the impact of noise on prediction.