Analyse de la performance de la méthode d'imputation de données manquantes missForest et application à des données environnementales

Abstract

Data acquisition and recording in the form of databases for later exploitation are routine operations in most fields (medicine, industrial production, education, environment, etc.). However, measurement, acquisition and/or recording processes may malfunction and cause data in the database to be missing. This missing data alters the subsequent analysis efficiency and, consequently, information and associated decision-making. Furthermore, because of the broad spectrum of activities that have an impact on natural environments, the databases collected and recorded in environmental matters are generally of a mixed nature (quantitative and qualitative). In this context, it becomes relevant to evaluate the performance of missing data processing methods considering this characteristic. In this study, missing data imputation methods were investigated and more specifically, the performance of the missForest method and its application to the problem of missing data in environment. Hence, a comparative study was carried out between missForest and two other imputation methods, Multivariate Imputation by Chained Equations (MICE) and K-nearest neighbors (KNN). This comparative analysis took into account 10 complete databases of various types (qualitative, quantitative and mixed data) specifically considering actual imputation error indicators and processing time. The application of the missForest method to the treatment performance database of Quebec’s wastewater treatment plants was then carried out as a case study of environmental data. The results of the comparative study revealed that in terms of imputation errors, missForest was the most efficient method for 9 out of 10 tested databases. The performance gap was more evident for mixed data imputations, missForest has reduced imputation errors up to 60 % in regard to the other methods. Concerning processing times, KNN was the fastest method for all of the databases when the missing data percentage was less than or equal to 30 %. The missForest processing times, although generally higher than those of KNN, tended to decrease with the increase in the percentage of missing data. The application of the missForest method to wastewater treatment plants data led to estimated errors systematically lower than 10 %. These results suggest that missForest is an imputation method that should be preferred when dealing with missing data in environment.

Date	26 Jun 2019
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Mathias Glaus (Supervisor) & Fausto Errico (Co-supervisor)