Human action recognition poses a challenge in computer vision due to various environmental factors and the need for extensive labelled data in traditional deep learning approaches. To address this, a semi-supervised learning framework is proposed that leverages feature-based clustering to utilize both labelled and unlabelled data effectively. The performance of SVM, random forest, decision tree, logistic regression, and XGBoost classifiers are evaluated in this framework. Our results show that logistic regression achieves the highest accuracy of 0.86, followed by SVM with 0.77, and XGBoost with 0.77. Random forest and decision tree classifiers exhibit lower accuracies of 0.74 and 0.3975, respectively. The findings demonstrate the potential of semi-supervised learning methods for improving human action recognition performance, particularly in scenarios with limited labelled data. The study lays a foundation for future research in semi-supervised learning for human action recognition, with implications for real-world applications.