Behavior Identification of Social Individuals from Sensor Data

Nowadays, biologists and social scientist are interested in understanding the fundamental properties and dynamics of a group of individuals. How, what, and why an organism acts in its life is a key component of this understanding. Individuals and groups of individuals are defined by their behaviors, thus the behavioral ecology field has focused on developing methods to observe and annotate these behaviors. These methods are capable of collecting data about multiple types of activities, from individual to social, and they can be applied to most of the social species in this world. Ethograms are the principal tool of behavioral data collection. Scientist collect behavioral annotations based on the behavioral structure contained in the ethograms. However, these methods require the presence of one or more people observing the animals at all times. In general, a single person or even a few people can not watch all the animals all the time. To overcome this problem, recently, thanks to the development of tracking technologies, biologists have started augmenting animals with sensors that are able to remotely collect data about the movement and the position of an organism. This new technology allows for collecting enormous amount of data, which in their raw format, are not human interpretable. It is necessary to translate these raw data into humanly understandable behaviors. In this work, I propose a solution to the problem of behavior identification of social indi viduals from sensors data. The solution is based on a machine learning framework able to infer behaviors at multiple levels of aggregation. The framework is based on the common structure of a time series analysis methods. First, temporal resolution selection is performed. Secondly, temporal feature extraction is performed and, finally, the actual classification is performed. The framework has been developed in order to exploit the underlying social structure of social individuals, furthermore, the ethogram is used to drive the entire classification process. The data considered in this work has been collected on Olive baboons in 2012, at the Mpala Research Centre in Kenya. In this case, biologists were interested both in individual and social group behaviors. The results have shown that the proposed method is able to achieve and obtain better results than the generic state-of-the-art methods, both in term of F1 and accuracy score. Although the proposed framework has been validated on social animals, the general structure and, in particular, the use of the behavioral hierarchy to drive the classification can be exported to a variety of domains, including humans.