Using supervised machine learning to quantify cleaning behaviour

Abstract

Cleaner fish engage in mutualistic interactions by removing ectoparasites from client species, a behaviour that has traditionally been quantified through labour-intensive manual video analysis. This method is not only time-consuming but also susceptible to human error and bias. In this study, we developed a semi-automated system to track and classify cleaning interactions between the cleaner wrasse (Labroides dimidiatus) and the powder blue tang (Acanthurus leucosternon) in a controlled three-dimensional laboratory setting. We employed DeepLabCut (DLC), a deep learning-based tool for markerless pose estimation, to track both fish species simultaneously. The resulting model reliably tracked both individuals with low error rates. Using the tracking data, we designed a classification algorithm that detected cleaning interactions with 90% accuracy. Although the algorithm misclassified approximately 15% of non-interactions as interactions, it successfully identified 25% of video content as containing interactions, thereby reducing the amount of footage requiring manual annotation by 75%. This approach significantly decreases human labour while maintaining high classification performance. Overall, our system represents a valuable step toward automating behavioural analysis in marine mutualisms and can serve as a foundation for broader applications in ethology and conservation research.