The number of the applications that analyze and evaluate human activities of daily living such as transport mode detection and activity recognition is increasing rapidly due to the require-ments in several fields such as transportation planning, elderly care and ambient assisted liv-ing. One of the drawbacks of these systems is their high battery consumption characteristics. In this study, we introduce a novel instance selection methodology that provides energy saving in testing process by reducing the amount of the training data while preserving the accuracy of the system. By their nature, daily living activities separate to several sub-classes within each class. The proposed method selects instances in an iterative cluster-based manner assorting with the characteristic structure of the daily activities. The success of the system is evaluat-ed by applying Decision Tree (J48) and k-Nearest Neighborhood (k-NN) algorithms to two different publicly available daily activity datasets. Obtained results show that the proposed instance selection algorithm based on sub-activity characteristics could achieve up to 11% im-provement of the classification results when 50% of the training instances are eliminated. With the help of this selection process, we built 4% to 57% smaller and 4% to 62% faster models for activity recognition on energy-constrained devices.