Class-Specific feature selection for enhanced cattle behavior recognition using wearable-based system

Monitoring and analyzing behaviors of individual cattle is critical for assessing health and performance, thereby enhancing production management in modern animal husbandry. Current behavior recognition systems typically rely on wearable sensors and standard pipelines that involve feature extraction followed by selection to build classification models. However, a key limitation of conventional feature selection methods is their use of a uniform feature subset across all behavioral categories. This approach fails to capture class-specific feature discriminative patterns and can lead to degraded classification performance as the number of behavior classes increases. To address this gap, this paper proposes a Class-Specific Feature Selection (CSFS) framework that restructures multi-class behavior recognition by decomposing the original C-class task into C distinct binary classification sub-tasks. This allows the selection of optimal feature subsets tailored to each behavioral category based on its unique characteristics. Final predictions are generated through an ensemble integration mechanism. Experimental comparisons and robustness evaluations demonstrate that the proposed framework achieves an average accuracy improvement of 21% over conventional feature selection methods ( p <0.05), a statistically significant enhancement that substantially boosts recognition reliability for precision livestock applications. The framework is compatible with arbitrary feature selection algorithms and machine learning classifiers, and exhibits superior cross-scenario robustness under noisy conditions and different sensor combinations. By providing deeper behavioral insights and enabling more accurate multi-behavior concurrent recognition, the proposed framework supports the development of data-driven precision grazing systems and contributes to sustainable livestock management practices.