Performance analysis of feature-based automated measurement of mouse social behavioral

Automated social behavior analysis in the mammalian animal has become an increasingly popular and attractive alternative to traditional manual human annotation with the advancement of machine learning and video tracking system for automatic detection. In this work, we study a framework of how different features perform on the different classifiers to analyze automatic mice behavior. We conducted experiments on the Caltech Resident-Intruder Mouse (CRIM13) dataset, which provides two types of features: trajectory features and spatio-temporal features. With this feature, we train Ada Boost and Random Decision Forest (Tree Bagger) classifiers to classify different mouse behaviors to show which features perform best on which classifier. The experimental result shows that the trajectory features are more informative and provide better accuracy than the widely used spatio-temporal features, and Ada Boost classifier shows better performance than the Tree Bagger on these features.