Towards Calibrated, Sharp, and Interpretable Probabilistic Prediction

Abstract

Predicting the future has long been a fundamental aspiration of humankind. In the current era of big data, the capacity to collect vast amounts of information presents a substantial challenge in effectively leveraging this data, given its increasing diversity, dynamism, and scale. In addition, data-driven methods, especially the evermore powerful deep learning models, are lacking in terms of interpretability and uncertainty quantification. This dissertation focuses on the theory and the practice that improve the calibration, sharpness, and interpretability in probabilistic forecasting, with a particular focus on applications in solar flare forecasting and active region evolution prediction. In terms of data, we devise methods that fuse historical data, combine multimodal data, and construct physically meaningful features, to unleash the potential of deep neural networks in solar flare forecasting. In terms of models, we apply modern methods to estimate the high-dimensional multiway covariance, and use them to predict the evolution of the high-dimensional multi-channel spatio-temporal process observed in solar active regions. In terms of post-processing, we develop statistical theories and efficient algorithms in binary recalibration problems with potential distribution shifts, and apply them in solar flare prediction regulated by solar cycles. In terms of post-hoc interpretation, we develop attribution tools for astrophysicists to identify the precursors of solar flares discovered by convolutional neural networks. With these theoretical and practical advances, we provide refreshing understanding of the statistical theory that underpins the general probabilistic prediction, and practical predictive methodology and techniques that utilize massive and complex data.