What is the purpose of Multi-Window Temporal Contrasting (MWTC) and how does it work?

Multi-Window Temporal Contrasting (MWTC) is used to maintain temporal consistency in multivariate time series learning at the individual sensor level. It summarizes past window features from one augmented view and contrasts them with future window features from the other view, so that predicted future windows align with their true counterparts while being distinguished from other window features. This helps preserve the natural temporal structure of the data and produces high-level sensor representations that also improve graph contrasting.

MWTC operates at the scale of each sensor across multiple time windows. Because later windows depend on earlier ones, it exploits these temporal dependencies to make learned representations robust to perturbations. For every sensor, an auto-regressive function compresses features from a set of past windows in one view into a summary vector. That summary is then used through nonlinear mappings to predict upcoming window features in the other augmented view. The resulting contrastive loss encourages the predicted future window features to match the actual future window features in that other view, while distinguishing them from other nonmatching window features. The process is run in both directions, from strong to weak views and from weak to strong views. According to the source, the summarization function is implemented as a transformer. The summary vectors produced by MWTC are also used as inputs to graph contrasting, so MWTC supports spatial consistency as well.