A Variational Bayesian Approach for Fast Adaptive Air Pollution Prediction

Abstract

Air pollution problem has been a worldwide environmental concern in recent years. Accurate air pollution prediction can effectively protect public health and help government decisions. However, strong instability and frequent pattern shift in air pollution data challenges the conventional time series prediction paradigm and attracts interests in adaptive prediction algorithm. Recent progress in deep learning community, such as attention mechanism and meta learning algorithm, both use handcrafted adaptive strategy and lack sufficient usage of supporting observation data. In this paper, we adopt a variational Bayesian approach to enable fast adaption ability for a given air pollution predictor, which can make better use of recent observation data and adaptively inference task-specific parameters to achieve better adaption performance. Specifically, without explicitly designing a heuristic adaptive procedure, we formulate the adaptive prediction as a maximizing conditional likelihood problem on a generative graphic model, where a variational approximation to the intractable likelihood is further derived for end-to-end training. Experiments on real-world air pollution datasets show significant improvements of the proposed method compared to previous works.