A Simple Optimization Strategy via Contrastive Loss for Recognizing Human Activity Using Wearable Sensors

The key toward sensor-based human activity recognition (SHAR) recently lies in how to learn more contextual representation from complex activities. In this work, we present a simple-yet-effective optimization strategy based on mutual information maximization to directly optimize the distance or similarity between samples and thus control the shape of the decision boundary for recognizing complex activities. Specifically, our optimization introduces contrastive loss into the HAR task to generalize the performance. The contrastive loss can maximize the mutual information between similar samples, where the positive selection is significant for learning decision boundaries. To this end, we also propose a positive selection strategy to avoid the inconsistency of the amplified in time and space when augmentation, which leverages the different representations of the same samples as positives and the corresponding contradiction pairs as negatives. Extensive experiments conducted on three benchmark SHAR datasets, i.e., PAMAP2, USC-HAD, and UNIMIB, demonstrate the superiority of our optimization strategy. It is worth mentioning that by applying our strategy, excellent results can be achieved just with a shallow ResNet encoder.