Behavior-specific binary machine learning models: Bout length of behavioral elements as biologically relevant parameter improves machine learning accuracy in analysis of dog behavior sequences

Machine learning methods are frequently used to detect behavioral and ecological data patterns. Even though these new mathematical methods are useful tools, the results are often ambivalent if we do not utilize biologically relevant parameters in the analyses. In our experiment, we analyzed whether the bout length of behavior elements could be a relevant parameter to determine the window length used by the machine learning method. We defined eight behavior elements and collected motion data with a smartwatch attached to the dog’s collar. The behavior sequences of 56 freely moving dogs from various breeds were analyzed by deploying a specific software (SensDog). The behavior recognition was based on binary classification that was evaluated with a Light Gradient Boosted Machine (LGBM) learning algorithm, a boosted decision-tree-based method with a 3-fold cross-validation. We used the sliding window technique during the signal processing, and we aimed at finding the best window size for the analysis of each behavior element to achieve the most effective settings. Our results showed that in the case of all behavior elements the best recognition with the highest AUC values was achieved when the window size corresponded to the median bout length of that particular behavior. In summary, the most effective strategy to improve significantly the accuracy of the recognition of behavioral elements is using behavior-specific parameters in the binary classification models, choosing behavior-specific window sizes (even when using the same ML model) and synchronizing the bout length of the behavior element with the time window length.