Enhancing Human Activity Recognition with Ensemble of Hybrid Feature Selection Techniques

Human activity recognition (HAR) has gained significant attention in recent years. However, benchmark HAR datasets such as UCI-HAR and UCI-HAPT extract various features from time, frequency, and wavelet domains to identify activities, there is still uncertainty about how to determine the most effective features for quickly and accurately recognizing the activities. Various feature selection algorithms have been developed to improve the recognition performance. In this work, we have made a comparative study of the effectiveness of ensemble technique to classify various human activities. We have proposed an ensemble of hybrid feature selection (EHFS) method, which considers the hybrid method of traditional feature selection to obtain an ensemble of feature sets. The features are then fed to various machine-learning algorithms and also the proposed ensemble classifier. Experiments has been performed on benchmark datasets i.e., UCI-HAR and UCI-HAPT, and the ensemble classifier outperforms all the other individual classifiers, achieving an accuracy of above 98% and 96% after hybrid feature selection for the UCI-HAR and UCI-HAPT datasets, respectively. Also, by considering a handful of features with the most relevant features, we can reach the maximal achievable accuracy considering all features. We have shown that even on reducing the features set by almost 99% (6 and 9 features out of 561), we can achieve an accuracy of 89% and 84% of the maximal achievable accuracy for UCI-HAR and UCI-HAPT datasets, respectively. Likewise, if we consider the ensemble feature set, we can reduce the more than 83% and 81% of the total dataset (i.e., 99 and 109 features out of 561), achieving near about maximal achievable accuracy considering all features for UCI-HAR and UCI-HAPT dataset respectively.