May, 2023

层次混合网络用于连贯的概率预测

TL;DRHierarchical Mixture Networks (HINT) is a model family for efficient and accurate forecasting, specialized on multivariate mixture with coherence via bootstrap reconciliation, and robustified against time series scale variations with normalized feature extraction and recomposition, achieving 8% sCRPS improved accuracy across five datasets compared to the current state-of-the-art, while conducting ablation studies on its components and investigating theoretical properties.