Abstract
Accelerometer-based livestock behavior classification depends on data representation strategies that influence predictive performance and computational requirements. This study compared raw acceleration signals (Raw), a complete set of 46 handcrafted features (Full), and a selected subset of 12 features (Selected) across five heterogeneous datasets involving cattle, sheep, and goats, with sampling frequencies ranging from 1 to 25 Hz. Random Forest, Support Vector Machine, k-Nearest Neighbors, and neural network architectures were evaluated using five-fold cross-validation with dataset-specific grouping procedures, with macro-F1 as the primary performance metric. Across 15 dataset–classifier combinations involving traditional algorithms, mean macro-F1 increased from 0.561 with Raw to 0.691 with Full and 0.670 with Selected. Feature selection reduced input dimensionality by 73.9%, with an average macro-F1 decrease of 0.021 relative to Full. However, the one-dimensional convolutional neural network achieved a macro-F1 of 0.835 using Raw acceleration signals in one of the evaluated datasets, highlighting the dataset-dependent effectiveness of different representation strategies. Feature-based representations also reduced inference time and model storage in several configurations, although computational benefits were not consistent across algorithms. These findings indicate that no single representation was universally optimal. Selecting appropriate accelerometer representations requires balancing predictive performance, behavioral characteristics, and computational requirements, particularly when developing resource-constrained livestock monitoring systems.