A new hybrid system combining active learning and particle swarm optimisation for medical data classification. (24th August 2021)

• Record Type: Journal Article
• Authors: Zemmal, Nawel; Azizi, Nabiha; Sellami, Mokhtar; Cheriguene, Soraya; Ziani, Amel
• Is Part Of: International journal of bio-inspired computation. Volume 18:Number 1(2021)
• Page Start: 59
• Page End: 68
• Publication Date: 2021-08-24
• Subjects: active learning -- uncertainty sampling strategy -- particle swarm optimisation -- PSO -- global optimisation problem -- instance weighting -- informativeness -- unlabeled data -- medical data; Biologically-inspired computing -- Periodicals; Computational biology -- Periodicals; 572.0285
• Languages: English
• ISSNs: 1758-0366

Abstract

With the increase of unlabeled data in medical datasets, the labelling process becomes a more costly task. Therefore, active learning provides a framework to reduce the amount the manual labour process by querying an expert for just the labels of particular instances, the choice of these instances to annotate is paramount. However, the traditional active learning techniques can be computationally expensive as they require to analyse, at each iteration, all unlabeled instances including those that are redundant and uninformative, thereby decreasing the system performance. To handle this issue, it is necessary to have a global optimisation algorithm that allows finding the best solution in a reasonable time. This paper proposes a novel framework combining active learning and particle swarm optimisation algorithm. A novel uncertainty-based strategy was designed and integrated into the PSO as an objective function. This new strategy allows finding the most informative instances by calculating an uncertainty score using instance weighting method. Experiments were performed on binary and multi-class classification problems using both balanced and unbalanced medical datasets. Experimental results show that the proposed uncertainty strategy outperforms its existing counterparts. It achieves performances comparable to supervised methods.