Abstract: Daily human activity recognition using sensor data can be a fundamental task for many real-world applications, such as home monitoring and assisted living. One of the challenges in human activity recognition is to distinguish activities that have infrequent occurrence and less distinctive patterns. We propose a hierarchical classifier to perform two-phase learning. In the first phase the classifier learns general features to recognise majority classes, and the second phase is to collect minority and subtle classes to identify fine difference between them. We compare our proposal with a collection of state-of-the-art classification techniques on four real-world third-party datasets that involve different types of object sensors and are collected in different environments and on different subjects and six imbalanced datasets from the UCI-Irvine Machine Learning repository. Our results demonstrate that our hierarchical classifier approach performs better than state-of-the-art techniques including both structure- and feature-based learning techniques. The key novelty of our approach is that we reduce the bias of the ensemble classifier by training it on a subspace of data, which allows identification of activities with subtle differences, and thus provides well-discriminating features.