Machine-learning models, cost matrices, and conservation-based reduction of selected landscape classification errors

Abstract

Use of statistical models developed with machine-learning algorithms is increasing in the ecological sciences, yet these disciplines have not capitalized on the ability to use cost matrices to selectively reduce classification errors that have highly detrimental consequences. Our aim was to promote such applications by demonstrating the process of using a cost matrix to decrease specific types of misclassification, explaining the importance of exploring the effectiveness of cost matrices for a given dataset, and encouraging use of cost matrices with machine-learning models in landscape-ecological and conservation contexts. Bird occurrence data, landscape and regional land-cover data, costs of false-positive and false-negative errors, and the C5.0 decision tree algorithm were used to train and test a binary classifier. Increasing the cost for false negatives tended to decrease the frequency of this error type while allowing for reasonable predictive performance for each class separately and both classes combined. Cost matrices are applicable to many different categorical response variables and spatial scales. We encourage landscape ecologists and planners to explore the effectiveness of cost matrices for their particular dataset and project goals, especially when conservation of biodiversity across broad spatial extents is at stake.