EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

3967. Mixture of Experts-based Coherent Probabilistic Forecasts for Hierarchically related Time Series

Contributed abstract in session MA-4: Recent Methodologies in Explainable AI (XAI) 1, stream Recent Advancements in AI .

Monday, 8:30-10:00
Room: 1001 (building: 202)

Authors (first author is the speaker)

1. Junyong Lee
Department of Industrial and Management Engineering, Korea University
2. Jun-Geol Baek
School of Industrial Management Engineering, Korea University

Abstract

Coherent probabilistic forecasting, where the goal is to forecast multivariate time series that have hierarchical aggregation is important for accurate decision making in real world application. Recent work has focused on end-to-end fashion that simultaneously learns from all time series in the hierarchy and comprise the reconciliation step in single trainable model. However, this approach often neglects that distinct advantages of several reconciliation methods, where is adept to identifying temporal characteristics. To address this gap, our study proposes a novel coherent probabilistic forecasting framework with mixture-of-experts (MoE). Employing the MoE framework, our approach effectively captures both regular and irregular temporal dependencies to integrate three diverse strategies: top-down, bottom-up and fully coherent. By introducing different experts, this model can generate forecasts that are probabilistically coherent as well as produce sophisticated forecasts that encompass diverse temporal dynamics. An empirical evaluation on real world hierarchical time series datasets demonstrates comparable advantages of the proposed approach.

Keywords

Status: accepted

Back to the list of papers