Optimal Activity Recognition Framework Based on Improvement of Regularized Neighborhood Component Analysis (RNCA)

This study aims to investigate the alternative model structure based on a feature selection algorithm on multiple features-framework of human activity recognition (HAR) via wearable sensor-based modality. Neighborhood component analysis (NCA) is a linear transformation that maximizes the accuracy of specified classification events used as the benchmark for the proposed algorithm. Also, the effect of different combinations of sensor configurations of two, three, and all four sensors on the performance of the developed model was studied. The effectiveness and shortcomings of best sensor configuration were highlighted. Results were compared between different sensor configurations and benchmark HAR dataset. To maximize the regularization of NCA, fine-tuning the algorithm to maximize relevance and minimize redundancy (MRMR) was proposed. Results demonstrated that RNCA-MRMR could establish an efficient algorithm that can satisfy the model validation tests with significant advantages over feature number and predictive accuracy at 93.5%, 93.7%, and 94.5% for two, three, and all four sensors respectively. Furthermore, the adaptability of RNCA-MRMR to different data characteristics has ensured an optimal and task-specific representation of the data. In essence, the combined strength of RNCA and MRMR provides a versatile and effective approach for extracting meaningful features and enhancing the overall performance of machine learning models.