Biclustering is an unsupervised machine-learning technique that simultaneously clusters rows and columns in a data matrix. Over the past two decades, the field of biclustering has emerged and grown significantly, and currently plays an essential role in various applications such as bioinformatics, text mining, and pattern recognition. However, finding significant biclusters in large-scale datasets is an NP-hard problem that can be formulated as an optimization problem. Therefore, metaheuristics have been applied to address biclustering problems due to their (i) ability to efficiently explore search spaces of complex optimization problems, (ii) capability to find solutions in reasonable computation time, and (iii) facility to adapt to different problem formulations, as they are considered general-purpose heuristic algorithms. Although several studies on biclustering approaches have been proposed, a comprehensive study using metaheuristics for bicluster analysis is missing. This work presents a survey of metaheuristic approaches to address the biclustering problem in various scientific applications. The review focuses on the underlying optimization methods and their main search components: representation, objective function, and variation operators. A specific discussion on single versus multi-objective approaches is presented. Finally, some emerging research directions are presented.
14 March 2023 in International Transactions in Operational Research
In this paper, we propose a method for solving a real‐world timetabling problem at Mandarine Academy. The primary motivation for this work is to provide an automated professional course scheduling tool to replace the time‐consuming task of manually creating timetables that are constantly incorrect. Following a review of both scientific literature and company requirements, a mathematical model of the problem is provided, which includes 18 constraints (hard/soft) and five objectives, two of which are competing. We test a handful of multi‐objective evolutionary algorithms (MOEA's) starting with the non‐dominated sorting genetic algorithm (NSGA II and NSGA III), the multi‐objective evolutionary algorithm based on decomposition (MOEA/D), the indicator‐based evolutionary algorithm and finally the strength Pareto evolutionary algorithm . Two custom genetic operators (mutation and crossover) are proposed and compared to conventional operators (PMX and swap mutation). To obtain elite configurations, a tuning phase involving all of the aforementioned algorithms is carried out. Experiments were divided by problem size, with three to five objectives tested. Experiments include the use of real‐world data from the company's catalog. This dataset was made available to the scientific community to serve as a testing ground for professional course scheduling, an underexploited field of scheduling. We discuss findings, including a comparison of each algorithm's performance using various metrics, as well as convergence graphs and population evolution.
Les données médicales souffrent de problèmes d'uniformisation ou d'incertitude, ce qui les rend difficilement utilisables directement par des logiciels médicaux, en particulier dans le cas du recrutement pour les essais cliniques. Dans cette thèse, nous proposons une approche permettant de palier la mauvaise qualité de ces données à l'aide de méthodes de classification supervisée. Nous nous intéresserons en particulier à 3 caractéristiques de ces données : asymétrie, incertitude et volumétrie. Nous proposons l'algorithme MOCA-I qui aborde ce problème combinatoire de classification partielle sur données asymétriques sous la forme d'un problème de recherche locale multi-objectif. Après avoir confirmé les apports de la modélisation multiobjectif dans ce contexte, nous calibrons MOCA-I et le comparons aux meilleurs algorithmes de classification de la littérature, sur des jeux de données réels et asymétriques de la littérature. Les ensembles de règles obtenus par MOCA-I sont statistiquement plus performants que ceux de la littérature, et 2 à 6 fois plus compacts. Pour les données ne présentant pas d'asymétrie, nous proposons l'algorithme MOCA, statistiquement équivalent à ceux de la littérature. Nous analysons ensuite l'impact de l'asymétrie sur le comportement de MOCA et MOCA-I, de manière théorique et expérimentale. Puis, nous proposons et évaluons différentes méthodes pour traiter les nombreuses solutions Pareto générées par MOCA-I, afin d'assister l'utilisateur dans le choix de la solution finale et réduire le phénomène de sur-apprentissage. Enfin, nous montrons comment le travail réalisé peut s'intégrer dans une solution logicielle.