Gaussian-SAINT enabling probabilistic prediction and interpretation for high uncertainty events

Abstract

When making decisions based on regression forecasts for high-uncertainty events, there is a high risk of a disadvantage if the predictions differ greatly from the actual outcome. In these cases, stochastic forecasting models can be used to make decisions such as avoiding such a risk according to variance (uncertainty). However, the problem with such models is that it is difficult to obtain an interpretation of the output process of the prediction outcome directly from the model. Attention mechanisms, on the other hand, allow interpretation of the output process of prediction outcomes directly from the model. Among them, SAINT achieves highly accurate forecasts using two types of attention mechanisms. However, SAINT is based on point estimation, which is insufficient for decision making for high-uncertainty events. In this study, we propose an extension of SAINT, Gaussian-SAINT, which enables distribution estimation and multifaceted interpretation of prediction outcomes directly from the model. In addition, we conduct evaluation experiments on the proposed method and discuss its interpretability and contribution to decision making to demonstrate the usefulness of Gaussian-SAINT.