In July 2026, the Brain‑inspired Algorithm and Model Research Group at the Guangdong Institute of Intelligent Science and Technology, in collaboration with Tsinghua University, the Center for Excellence in Brain Science and Intelligence Technology (Institute of Neuroscience) of the Chinese Academy of Sciences, and other institutions, published a paper entitled “A Benchmark Dataset for Rat Social and Aggressive Behavior Classification” in Scientific Data, an international journal under the Nature Portfolio. Concurrently, they publicly released a video dataset for the study of rat social behaviour.

The study, based on the classic Resident–Intruder Paradigm, collected free‑interaction videos of rats and established a standardised annotation system covering three coarse‑grained behavioural categories—non‑social, social, and aggressive—with aggressive behaviour further subdivided into seven ethologically defined subclasses. The work also provides a reproducible analysis pipeline from pose estimation to behaviour classification, along with a unified evaluation benchmark, offering an open resource for objective and large‑scale quantitative analysis of social behaviour.

Social behaviour between conspecific individuals is a core research topic in behavioural neuroscience and ethology, yet its neural mechanisms remain poorly understood. A major obstacle is that social interactions are inherently highly dynamic and difficult to quantify: rapid behavioural transitions, frequent physical contacts, and subtle postural changes pose substantial challenges for automated behaviour recognition, and a large proportion of studies still rely on time‑consuming, expert‑based frame‑by‑frame manual annotation. Meanwhile, the field has long lacked standardised datasets, transparent annotation protocols, and reproducible evaluation benchmarks, making cross‑study comparisons difficult; early methods relying on body markers or wearable devices may also interfere with natural behaviour. Establishing an open, transparent, and reproducible quantitative framework for social behaviour without disrupting natural actions is therefore a critical step towards advancing the field.

Figure 1. Overall workflow of the study: behaviour video acquisition-keypoint tracking-feature computation-behaviour classification.

To address these challenges, this study carried out systematic work at three levels: dataset construction, analytical pipeline, and benchmarking.

Data collection. The team adopted the Resident–Intruder test paradigm and collected 58 top‑view videos of rats freely interacting, each continuously recorded for 10 minutes at 20 frames per second and 1080p resolution, yielding approximately 12,000 frames per video. Behavioural annotation followed a single‑blind procedure: videos were first randomised and re‑numbered, then independent annotators, unaware of animal identities and experimental conditions, labelled the start frame, end frame, and behaviour category for each event. Labels were first assigned to three coarse categories—non‑social, social (affiliative), and aggressive—and aggressive episodes were further subdivided into seven fine‑grained subclasses: chase, fight, lateral threat, bite, upright posture, pin, and freeze. Consistency tests yielded Cohen’s κ values of 0.931 for coarse‑grained and 0.887 for fine‑grained annotations, both in the “almost perfect agreement” range.

Figure 2. Coarse‑grained behavioural categories and fine‑grained subcategories of aggressive behaviour.

Analytical pipeline. The team provided a marker‑free, extensible pose‑driven workflow as a reference implementation. This pipeline uses DeepLabCut (DLC) to assign distinct identities to the two interacting individuals and track three anatomically stable landmarks visible in the top‑view perspective—head, body centre, and tail base. The resulting keypoint trajectories are then transformed into structured behavioural features: static features include keypoint coordinates, pairwise distances, and triplet angles; dynamic features capture temporal changes in posture and movement; and interaction features describing the spatial relationship between the two individuals are also introduced. This yields three progressive feature representations of 8, 14, and 26 dimensions. This representation uniformly maps complex social videos into quantitative features directly usable for classification, avoiding interference with natural behaviour caused by physical markers and facilitating transfer across different experimental setups.

Figure 3. Keypoint tracking results based on DeepLabCut and example keypoint coordinate trajectories.

Evaluation benchmark. Under a unified feature representation and evaluation protocol, the team systematically compared six representative models—CNN, LSTM, GMM, LightGBM, Random Forest, and XGBoost—using video‑wise data splitting to prevent information leakage and repeated experiments with five random seeds. Results showed that coarse‑grained social behaviour classification achieved substantially higher overall performance than fine‑grained aggressive behaviour recognition; tree‑based models and deep models performed comparably in prediction accuracy, but tree models offered clear advantages in training efficiency (LightGBM: 30.2 s, XGBoost: 54.6 s, versus LSTM: 887.7 s and CNN: 180.2 s). Notable performance discrepancies across behaviour categories were observed, with aggressive behaviours being the most difficult to identify and exhibiting the highest variability. Coefficients of variation across repeated runs were generally low, indicating good reproducibility of the benchmark results.

This work integrates raw videos, manual annotations, keypoint trajectories, and derived features into a structured open resource, providing a practical foundation for developing, evaluating, and comparing automated analysis methods for rat social and aggressive behaviour, and offering a scalable behavioural quantification tool for further investigation into the neural underpinnings of social behaviour.

Xutian Chen and Guangyu Li from the Guangdong Institute of Intelligent Science and Technology, and Zihan Zhang from the Center for Excellence in Brain Science and Intelligence Technology (Institute of Neuroscience), Chinese Academy of Sciences, are the co‑first authors of this paper. Dr. Mingkun Xu, a Junior Researcher at the Guangdong Institute of Intelligent Science and Technology, and Dr. Qianqian Shi from the Center for Brain‑Inspired Computing Research (CBICR), Department of Precision Instrument, Tsinghua University, are the co‑corresponding authors. Professor Zuoren Wang from the Institute of Neuroscience, Chinese Academy of Sciences, participated in and supervised the work. This research was supported by the High‑level Talent Innovation Team Project of the Guangdong‑Macao In‑depth Cooperation Zone in Hengqin, the National Natural Science Foundation of China, and other funding sources.

Project website:

https://blossom0913.github.io/Mouse-Behavior-Classifier-Train/

Paper link:

https://doi.org/10.1038/s41597-026-07888-8

Dataset link:

https://doi.org/10.6084/m9.figshare.30393298.v4

Code link:

https://doi.org/10.5281/zenodo.18299934