A Benchmark Dataset for Rat Social and Aggressive Behavior Classification

Area of Science:

• Behavioral Neuroscience
• Systems Neuroscience
• Computational Neuroscience

Background:

• Understanding the neural basis of social behavior requires reliable quantification of animal interactions.
• Limited availability of standardized datasets and benchmarks hinders progress in automated social behavior classification.

Purpose of the Study:

• To present a curated video dataset of rat social interactions using the resident-intruder paradigm.
• To develop a reproducible pose-based workflow for automated analysis of social behaviors.
• To benchmark machine learning models for social behavior classification and aggression recognition.

Main Methods:

• Collected and annotated a video dataset of rat social interactions, including fine-grained aggression subtypes.
• Developed a pose-based workflow using DeepLabCut to extract movement and spatial features.
• Benchmarked traditional and deep learning models using a unified evaluation protocol.

Main Results:

• Provided reference performance benchmarks for coarse-grained social behavior classification and fine-grained aggression recognition.
• Documented performance differences, training efficiency, and model complexity across evaluated models.
• Established a reproducible workflow for analyzing complex, contact-rich social interactions.

Conclusions:

• The curated dataset and workflow serve as an open resource for developing and comparing automated methods for rat social interaction analysis.
• Facilitates reproducible research in behavioral and systems neuroscience by standardizing social behavior quantification.
• Advances the field by providing tools for objective and scalable analysis of social dynamics.