Predicting Turning Points in Air Quality: A Dual-Guided Denoising Teacher-Student Learning Approach

The increasing application of IoT technology has facilitated the widespread deployment of air quality stations, making accurate air quality prediction increasingly essential for ensuring the stability of our daily lives and work environments. In this task, predicting abrupt changes or turning points is of utmost significance, as it helps to mitigate risk and prevent losses. However, the problems of future information deficiency and irrelevant sequence noise mislead the result to suboptimal performance. To overcome this problem, we propose a Dual-guided Denoising Teacher-Student learning approach (named TS-D \(^2\) ) to predict turning points in air quality. Since the complete distribution of turning points encompasses past and future data, we design a teacher network that integrates future data to guide the student network’s learning process that relies solely on the historical sequence. To only absorb credibly relevant features, the teacher network adds a Gumbel-enhanced denoising strategy for optimizing sequence. Based on this well-trained teacher, our student network employs a dual-guided learning mechanism to adaptively excavate important and salient knowledge for training and optimization purposes. Finally, we validate the effectiveness of our proposed TS-D \(^2\)  through extensive experiments conducted on three real-world datasets. The results demonstrate that our approach outperforms existing state-of-the-art baselines.