Multiclass classification methods in ecology. (February 2018)

• Record Type: Journal Article
• Authors: Bourel, M.; Segura, A.M.
• Is Part Of: Ecological indicators. Volume 85(2018)
• Page Start: 1012
• Page End: 1021
• Publication Date: 2018-02
• Subjects: Machine learning -- Ensemble methods -- Prediction -- Classification rules -- Phytoplankton trait-based approach; Environmental monitoring -- Periodicals; Environmental management -- Periodicals; Environmental impact analysis -- Periodicals; Environmental risk assessment -- Periodicals; Sustainable development -- Periodicals; 333.71405
• DOI: 10.1016/j.ecolind.2017.11.031
• Languages: English
• ISSNs: 1470-160X

Abstract

Abstract: Multiclass classification refers to the construction of a model able to classify a response variable that can take more than two classes. Most ecological indices are naturally multiclass ( e.g. water quality index: bad, regular, good) and the generation of models able to predict the output class in novel situations is required. In this study, we introduce seven representative multiclass classification techniques, classic and more recent, their rationale, advantages, disadvantages and a practical R code to implement them. These methods are: (1) Linear discriminant analysis (LDA), (2) a consensus of binomial logistic regression (CLR), (3) multinomial regression (MNR) and (4) support vector machine (SVM), (5) Classification and Regression Trees (CART), (6) Random Forest (RF) and (7) Stagewise Additive Modelling using a Multi-class Exponential (SAMME) loss function. We showed their implementation under simulated and a real data set to classify phytoplankton organisms into morphology-based functional groups. Results suggest that the nature of the data ( i.e. linear vs non-linear) influence the predictive ability of multi-class classification models. Real phytoplankton data was accurately classified (error < 0.05) by RF, SAMME and CLR, while SVM and CART were close to nominal 0.05 and LDA performed worst, with the higher error rate ( ca. 0.7). The expected behaviour of the response variable should be considered when choosing a model for multiclass classification. The use of the generalization error allows to objectively rank among competing models. We showed the differences in interpretability among models, which is critical to decipher causal relationships among variables or to design management plans. We hope this article contributes to increase the use of these techniques, some of them flexible and distribution-free, and to improve the quality of multiclass classification models applied to ecological problems.