Association Rule-based Soft Labeled Human Activity Recognition Using Wearable Sensors

Human activity recognition (HAR) with wearable sensors has been widely used in many human-centric applications, due to the prevalent of various IoT devices and low-cost wearable sensors nowadays. Typically, training a reliable HAR system requires activity labels to be both completely and precisely annotated. While, high-quality annotations are always quite difficult to acquire for users, even it is possible to collect a large number of activity data in practice. In this paper, an accurate and interpretable HAR model is constructed for achieving favorable recognition performance under the circumstance of limited activity labels. The essence of our proposal is to assign soft labels for those available unlabeled data by using the association discovery and hierarchical clustering techniques, which to some degree reduces the accuracy loss caused by the lack of activity labels in training set. Moreover, a soft label-based evidential association rule learning procedure is conducted for learning the relations between instance features and activity classes. Finally, all the generated association rules from available labeled or unlabeled data are fused for the belief classification of unknown input instances. Experiments on the public UCI smartphone and self-collected activity datasets demonstrate the feasibility and effectiveness of the proposal, in which the classification accuracy of the proposal improves at most 5% and 11% if compared with the classical (association) rule-based models CBA and BRBCS.