Use of Domain Labels during Pre-Training for Domain-Independent WiFi-CSI Gesture Recognition

Abstract

To minimize dependency on the availability of data labels, some WiFi-CSI based-gesture recognition solutions utilize an unsupervised representation learning phase prior to fine-tuning downstream task classifiers. In this case, however, the overall performance of the solution is negatively affected by domain factors present in the WiFi-CSI data used by the pre-training models. To reduce this negative effect, we propose an integration of the adversarial domain classifier in the pre-training phase. We consider this as an effective step towards automatic domain discovery during pre-training. We also experiment with multi-class and label versions of domain classification to improve situations, in which integrating a multi-class and single label-based domain classifier during pre-training fails to reduce the negative impact domain factors have on overall solution performance. For our extensive random and leave-out domain factor cross-validation experiments, we utilise (i) an end-to-end and unsupervised representation learning baseline, (ii) integration of both single- and multi-label domain classification, and (iii) so-called domain-aware versions of the aformentioned unsupervised representation learning baseline in (i) with two different datasets, i.e., Widar3 and SignFi. We also consider an input sample type that generalizes, in terms of overall solution performance, to both aforementioned datasets. Experiment results with the Widar3 dataset indicate that multi-label domain classification reduces domain shift in position (1.2% mean metric improvement and 0.5% variance increase) and orientation (0.4% mean metric improvement and 1.0% variance decrease) in domain factor leave-out cross-validation experiments. The results also indicate that domain shift reduction, when considering single- or multi-label domain classification during pre-training, is negatively impacted when a large proportion of negative view combinations contain views that originate from different domains within a substantial amount of mini-batches considered during pre-training. This is caused by the view contrastive loss repelling the aforementioned negative view combinations, eventually causing more domain shift in the intermediate feature space of the overall solution.