Correcting a bias in the computation of behavioural time budgets that are based on supervised learning
Abstract: Supervised learning of behavioural modes from body acceleration data has become a widely used research tool in Behavioural Ecology over the past decade. One of the primary usages of this tool is to estimate behavioural time budgets from the distribution of behaviours as predicted by the model. These serve as the key parameters to test predictions about the variation in animal behaviour. In this paper we show that the widespread computation of behavioural time budgets is biased, due to ignoring the classification model confusion probabilities. Next, we introduce the confusion matrix correction for time budgets—a simple correction method for adjusting the computed time budgets based on the model's confusion matrix. Finally, we show that the proposed correction is able to eliminate the bias, both theoretically and empirically in a series of data simulations on body acceleration data of a fossorial rodent species (Damaraland mole‐rat Fukomys damarensis). Our paper provides a simple implementation of the confusion matrix correction for time budgets, and we encourage researchers to use it to improve accuracy of behavioural time budget calculations.
Online Publication Date
Horizon 2020 Framework Programme (742808)