Correcting a bias in the computation of behavioral time budgets that are based on supervised learning

Abstract

1. Supervised learning of behavioral modes from body-acceleration data has become a widely used research tool in Behavioral Ecology over the past decade. One of the primary usages of this tool is to estimate behavioral time budgets from the distribution of behaviors as predicted by the model. These serve as the key parameters to test predictions about the variation in animal behavior. In this paper we show that the widespread computation of behavioral time budgets is biased, due to ignoring the classification model confusion probabilities. 2. 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. 3. 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). 4. 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 behavioral time budget calculations.