Impact of proximity definitions and sampling rates on social networks in pigs based on tracking using computer vision

In farm animals, social network analysis has become a popular approach to explore preferential associations. This study investigated how different spatial proximity definitions and sampling rates affect social networks based on proximity using computer vision. Video data collected over three days in 21 pens (6 pigs/pen), either enriched or barren, were analyzed using a tracking-by-detection method based on bounding boxes. Networks were constructed with five different definitions of proximity: (1) distance between centroids of bounding boxes < 50 cm, (2) occurrence of overlap of surfaces of bounding boxes, (3) surface overlap of bounding boxes > 20%, (4) a combination of (1) and (3), and (5) the harmonic mean of the distance between the two individuals. For each proximity definition, networks built with downsampled data were compared to a network built with 0.5 frames per second. The network metric degree centrality was less affected by proximity definitions compared to eigenvector centrality and clustering coefficient. To maintain high correlations with the complete network (r > 0.90), downsampling should not go beyond 1 frame every 6 min. This work shows how computer vision data can be used for social network analysis in livestock with limited space and choice of social environment, and how metrics depend on proximity definitions and sampling rates.