Unsupervised feature selection for interpretable classification in behavioral assessment of children

Abstract

In this paper, we consider a data set taken from the administration of the Behavior Assessment\ud System for Children test to 157 subjects, and we approach the problem of clustering and classify\ud the subjects in an interpretable fashion. Because the Behavior Assessment System for Children test\ud is originally composed of 149 questions (152 in the particular version used for this experiment), we\ud first propose a feature selection wrapper model composed by a multi‐objective evolutionary\ud algorithm, the iterative clustering method expectation–maximization, and the classifier C4.5 for\ud the unsupervised feature selection towards the classification of the data with two objectives:\ud maximizing the likelihood of the clustering model and maximizing the accuracy of the obtained classifier.\ud We propose a methodology to integrate feature selection for unsupervised classification,\ud model evaluation, decision‐making (to choose the most satisfactory model according to an a\ud posteriori process in a multi‐objective context), and testing. The selected data set that is the result\ud of this process, where each instance is labeled with its class, is then used for supervised learning via\ud both C4.5 and a novel evolutionary computation‐based fuzzy classifier to obtain interpretable rules.\ud We discuss and compare the behavior of two different evolutionary algorithms (ENORA (Evolutionary\ud NOn‐dominated Radial slots based Algorithm) and NSGA‐II (Non‐dominated Sorted\ud Genetic Algorithm)) at different levels: as search strategies for feature selection, as search strategies\ud for fuzzy classification, and in terms of quality of the results. It turns out that ENORA behaves\ud better in terms of quality of the result in the feature selection phase (obtaining a selection that\ud shows higher accuracy under C4.5 after cross‐validation), and again in the fuzzy classification\ud phase, from both points of view: hypervolume evolution and interpretability of results. During\ud the entire process, the solutions are validated by the psychologists who collected the data

Temporal Aspects in Air Quality Modeling—A Case Study in Wrocław

“…Another wrapper method is proposed in Shi et al, 57 while in García-Nieto et al, 58 2 wrapper methods with 3 and 2 objectives, respectively, applied to cancer diagnosis are compared. Finally, very recent examples of multiobjective FS systems can be found in Jiménez et al 59,60 To the best of our knowledge, the only attempt to use multiobjective optimization process in modeling air quality has been used for monitoring system planning in Sarigiannis and Saisana. 61…”

Insights from a text mining survey on Expert Systems research from 2000 to 2016

“…Another domain that has recently been subjected to highly innovative research practices is health, with distinct examples of applications such as nutrition (Espín, Hurtado, & Noguera, 2016), patient care (James, Calderon, & Cook, 2017), depression (Chattopadhyay, 2014), and cancer (Acharya, Ng, Sree, Chua, & Chattopadhyay, 2014). Psychology and social sciences have also benefited from ESs' implementations (e.g., Jiménez, Jódar, Martín, Sánchez, & Sciavicco, 2016). Such diversity shows the broad applicability of ESs, providing solid grounded evidence of its relevance.…”

Multi-Objective Evolutionary Rule-Based Classification with Categorical Data

“…We decided to solve this problem by applying multi-objective evolutionary algorithms ( MOEA ) [ 15 , 16 ] as meta-heuristics, and, in particular, two known algorithms: NSGA-II [ 15 ] and ENORA [ 17 ]. They are both state-of-the-art evolutionary algorithms which have been applied, and compared, on several occasions [ 18 , 19 , 20 ]. NSGA-II is very well-known and has the advantage of being available in many implementations, while ENORA generally has a higher performance.…”