Semantic-Discriminative Mixup for Generalizable Sensor-based Cross-domain Activity Recognition

It is expensive and time-consuming to collect sufficient labeled data build human activity recognition (HAR) models. Training on existing often makes the model biased towards distribution of training data, thus might perform terribly test with different distributions. Although efforts transfer learning domain adaptation try solve above problem, they still need access unlabeled target domain, which may not be possible in real scenarios. Few works pay attention a that can generalize well unseen domains for HAR. In this paper, we propose novel method called Semantic-Discriminative Mixup (SDMix) generalizable cross-domain Firstly, introduce semantic-aware considers semantic ranges overcome inconsistency brought by differences. Secondly, large margin loss enhance discrimination prevent misclassification noisy virtual labels. Comprehensive generalization experiments five public datasets demonstrate our SDMix substantially outperforms state-of-the-art approaches 6% average accuracy improvement cross-person, cross-dataset, cross-position