Univariate Time Series Forecasting via Interactive Learning

For time series forecasting tasks, it is necessary to capture the temporal dependencies from observed variables. Although many deep learning models have gained good performance, they still lack an effective modeling of temporal dependencies. Additionally, statistical features of time series often change over time, resulting in distribution shift issues. This is also one of the main challenges for time series forecasting. In this paper, we propose a module called Interactive Temporal-spatial Attention (ITSA), which combines interactive convolution and attention mechanism to effectively model the dependence between time and suppress the distribution shift problem. First, the time series is normalized and decomposed into trend and seasonal components. We then use an interactive learning strategy to extract the temporal dependencies of observed values at different data resolutions. Next, a normalized temporal-spatial attention mechanism is employed to capture the temporal-spatial features of the time series to prevent information loss. Finally, the true distribution is obtained by inverting the normalized data to achieve the purpose of suppressing the distribution shift. We employ a hierarchical way to stack the proposed ITSA, namely HITSA, to complete the forecasting task. The experimental results show that the model has good predictive performance in datasets of electricity and MOOC, and is significantly superior to other baseline methods, which indicates that the proposed ITSA can extract representative features from time series.