SimPM: A Simple Patch Masking Contrastive Learning Framework for Time Series Forecasting

Time series forecasting plays a critical role in numerous practical
industries, where effectively learning and extracting meaningful representations has always been a significant and challenging problem. Although contrastive learning methods have shown outstanding ability in
learning meaningful representations in computer vision and natural language
processing domains, their performance in time series forecasting
tasks is weaker. This weakness can mainly be attributed to their failure
to fully consider the characteristics of time series data, leading to information loss.
Specifically, existing data augmentation strategies primarily
operate at the timestamp level, which cannot fully exploit and utilize
local semantic information. Moreover, previous research has not taken
into account the sharing of information between independent channels
when dealing with inter-channel information. This limitation, to some
extent, restricts the integrity of the learned representations. To address
these issues, we propose a new method called SimPM, a simple patch
masking contrastive learning framework for time series forecasting that
effectively mitigates information loss. In our experiments on seven benchmark
time series forecasting datasets, SimPM demonstrates competitive
performance compared to existing contrastive learning methods.