Abstract:
|
In supervised classification, the aim is either to allocate data to a set of pre-defined groups of interest or to discover sets of patterns, using a pre-classified dataset. For systems with temporal dependence, the goal is to correctly assign different segments of data to a finite set of groups that reflect important dynamics of interest, taking into account that observations near each other (in time) are likely to correspond to the same group. Animal accelerometer data is typically collected at a fine temporal resolution, leading to high temporal dependence of the observation process. As the temporal dependence provides valuable information for identification of behavioral patterns, we propose the use of the hidden semi-Markov model with an autoregressive structure for the observation process. In this manner, we can explicitly account for the temporal dependence in both the behavior and observation process and evaluate the prediction error and choice of loss function suitable for data that exhibits strong temporal dependence. We present a real data example for the classification of accelerometer data of Merino sheep into 5 behaviors. Inference is conducted in a Bayesian framework.
|