vassi – verifiable, automated scoring of social interactions in animal groups

Behavioral biologists, from neuroscientists to ethologists, rely on observation and scoring of behavior. In the past decade, numerous methods have emerged to automate this scoring through machine learning approaches. Yet, these methods are typically specified towards laboratory settings with only two animals, or employed in cases with well-separated behavioral categories. Here, we introduce the vassi Python package, focusing on supervised classification of directed social interactions and cases in which continuous variation in behavior means categories are less distinct. Our package is broadly applicable across species and social settings, including single individuals, pairs and groups, and implements a validation tool to separate behavioral edge cases. vassi has comparable performance to existing approaches on a behavioral classification benchmark, the CALMS21 mouse resident-intruder dataset, and we demonstrate its applicability on a novel, more naturalistic and complex dataset of cichlid fish groups. Our approach highlights future challenges in extending supervised behavioral classification to more naturalistic settings, and offers a methodological framework to overcome these challenges.