Inferring Interaction Network from Sensor Data

Being able to observe how animals interact among themselves has always been a crucial requirement for behavioral scientists who study social species. Physically watching them to see when interactions occur is extremely time-consuming, results in many missed observations and it becomes extremely difficult to do for a very extended period of time. This is especially true for animals who live in large territories and for animals who behave differently when nearby human beings. To overcome these problems, scientists are accustomed to use several different kinds of sensors which are usually attached to the target animals to record some kind of raw data which, once collected, is used to analyze the behavior of their hosts. The sensors have to be as least invasive as possible for the animal to behave as if it wasn’t wearing them. This implies that sensors have to be light with respect to the animal, they must not emit a considerable amount of heat and the wavelength they use must not be perceptible by the animals near them. The most popular type of data extracted from sensors attached to animals is Global Posi- tioning System (GPS) data. GPS data is very easy to extract and it can be used to efficiently track the positions of entire groups of animals. However, when we are interested in pairwise interactions between animals and not in their positions, GPS data is not very reliable either because of its low accuracy and because it might not be in line of sight with the satellites it relies on. In this study, we introduce synthesized sensor data based on a type of non-invasive short range proximity sensor in order to understand whether some animals interact among themselves at a given time. It is essential to label the data in order to be aware of the interactions as they occur during the traning phase. The sensors whose data we are interested in synthesizing can be used on animals that are capable of wearing them due to weight or size constraints and as long as the sensor range contains the interaction range of the animals we are interested in. The models we analyze in this work don’t assume either that all of the animals behave uniformly amongst each other and also the independence amongst interactions. We present a framework to synthesize proximity, location and speed data extracted from sensors with several different configurations and we present models to infer animal interactions. Finally, we evaluate the methodology we use by proposing a case study where we perform different analyses to understand when our models are fit for the inference task, what are the most critical parameters impacting their performances and when we should start assuming independence conditions to simplify the task.