Unsupervised segmentation of the contextual structure inherent in goal-directed behavior in non-human primates

Conventional behavioral ethograms rely on experts’ estimation and labeling of specific time-series semantic structures inherent in human and animal behavior. Computational approach is expected to significantly improve the process of behavioral analysis and provide objective measure. In this study we segmented a continuous behavioral sequence in feeding behavior of common marmosets into repetitively observed units using a state-space model that incorporates unsupervised machine learning classification and nonlinear regression. We focused on the behavior because it supposed to have a specific semantic structure in the flow of actions to achieve the goal. A total of 1000-sec time series of feeding behavior of four marmosets was clearly divided into a sequence of 11 different segments. Characteristic segments appeared before and after the feeding timing, and the transition probabilities of these segments allowed us to classify three different feeding patterns. These results suggest that the proposed computational method has the potential to classify and quantitatively compare behavioral ethograms without subjective observation. It has opened the path to understand complex and ambiguous behavioral repertories in social communication and interaction, which are difficult to analyze using conventional methods, as statistical generative models.The 81st Annual meeting of Japanese society for animal psycholog