Automated face recognition using deep neural networks produces robust primate social networks and sociality measures

Longitudinal video archives of behaviour are crucial for examining how sociality shifts over the lifespan in wild animals. New approaches adopting computer vision technology hold serious potential to capture interactions and associations between individuals in video at large scale; however, such approaches need a priori validation, as methods of sampling and defining edges for social networks can substantially impact results.Here, we apply a deep learning face recognition model to generate association networks of wild chimpanzees using 17 years of a video archive from Bossou, Guinea. Using 7 million detections from 100 h of video footage, we examined how varying the size of fixed temporal windows (i.e. aggregation rates) for defining edges impact individual-level gregariousness scores.The highest and lowest aggregation rates produced divergent values, indicating that different rates of aggregation capture different association patterns. To avoid any potential bias from false positives and negatives from automated detection, an intermediate aggregation rate should be used to reduce error across multiple variables. Individual-level network-derived traits were highly repeatable, indicating strong inter-individual variation in association patterns across years and highlighting the reliability of the method to capture consistent individual-level patterns of sociality over time. We found no reliable effects of age and sex on social behaviour and despite a significant drop in population size over the study period, individual estimates of gregariousness remained stable over time.We believe that our automated framework will be of broad utility to ethology and conservation, enabling the investigation of animal social behaviour from video footage at large scale, low cost and high reproducibility. We explore the implications of our findings for understanding variation in sociality patterns in wild ape populations. Furthermore, we examine the trade-offs involved in using face recognition technology to generate social networks and sociality measures. Finally, we outline the steps for the broader deployment of this technology for analysis of large-scale datasets in ecology and evolution.