One of the important areas of machine intelligence research today is human activity recognition (HAR), with the goal of automatically identifying human activities from various types of sensor data. Most of the existing human activity recognition methods use hand-crafted features and labelled data, but these methods fail to identify new activities not defined in the training dataset. As human activities are numerous and executed in various ways, it is challenging to obtain enough labelled data to train a model to recognize the different activities. In this paper, the performance of five different supervised learning algorithms on the human activity recognition task with skeleton-based features has been evaluated, using five publicly available datasets and an experimental dataset. Accuracies of above 90% are achievable on datasets with a limited number of samples using commonly available classification algorithms and simple skeleton-based features. Subsequently, the same feature sets are used on unsupervised learning methods for an unsupervised clustering task. Using the unsupervised learning algorithms, an average of 74% f1-score on the publicly available CAD60 dataset and 61% f1-score on the experimental dataset, are obtained. These results demonstrate the effectiveness of simple skeleton-based features, coupled with common supervised and unsupervised learning algorithms in human activity recognition tasks.