A probabilistic framework for behavioral identification from animal-borne accelerometers

•Probabilistic framework for interpreting uncertainty- the probability profile.•Probability profiles describe behavioral variation predicted from machine learning.•Variation inherent in behavior measured with accelerometers incorporated into model.•Show creation of profiles with a random forest and clustered self-organizing map.•Used a matrix-based validation (L1-norm) to evaluate how well the models predict. Many studies of animal distributions use habitat and climactic variables to explain patterns of observed space use. However, without behavioral information, we can only speculate as to why and how these characteristics are important to species persistence. Animal-borne accelerometer and magnetometer data loggers can be used to detect behaviors and when coupled with telemetry improve our understanding of animal space use and habitat requirements. However, these loggers collect tremendous quantities of data requiring automated machine learning techniques to identify patterns in the data. Supervised machine learning requires a set of training signals with known behaviors to train the model to identify the unique signal characteristics associated with each behavior. In contrast, unsupervised approaches aggregate unlabeled signals into groups based purely on signal similarity but, without additional information, do not identify specific behaviors. In this paper, we propose a probabilistic framework for interpreting uncertainty in machine learning techniques—the probability profile—and demonstrate how to post hoc identify behaviors within signal groups. We assess model performance using a matrix-based measure of dissimilarity. We used a Random Forest (RF) and a clustered self-organizing map (CSOM) for comparison and demonstrate the use of a behavioral profile for each using a data set of high-frequency accelerometer and magnetometer data collected from 7 captive wild pigs (Sus scrofa) moving in a 1 ha outdoor enclosure. We found that the RF had more discrimination than the CSOM which had fewer clusters associated with high probabilities of a single behavior (>50%). The leave-p-out cross validation statistic of the probability matrix (L1¯) indicated that there was an average maximum dissimilarity of 20% and 65% between the training and test data sets for the RF and CSOM methods, respectively. Using a probability profile to describe groups predicted from machine learning allows the variation and error inherent in behavioral prediction to be incorporated direc