Robust Multi-Modal Forecasting: Integrating Static and Dynamic Features

TL;DR

The paper addresses robust, interpretable forecasting for multi-modal time series in high-stakes domains by extending TIMEVIEW to jointly model static features and dynamic time-series inputs. It introduces a trend-property encoding for exogenous dynamic features and uses a contrastive loss to align static and dynamic latent representations, aiming to mitigate sensitivity to measurement noise. Empirical results on synthetic datasets show that encoding trends and properties improves predictive accuracy over raw dynamic inputs while preserving interpretability and enabling counterfactual reasoning about temporal changes. This work advances transparent, robust forecasting by enabling cohesive integration of static and dynamic modalities with bi-level interpretability and applicability to healthcare settings.

Abstract

Time series forecasting plays a crucial role in various applications, particularly in healthcare, where accurate predictions of future health trajectories can significantly impact clinical decision-making. Ensuring transparency and explainability of the models responsible for these tasks is essential for their adoption in critical settings. Recent work has explored a top-down approach to bi-level transparency, focusing on understanding trends and properties of predicted time series using static features. In this work, we extend this framework by incorporating exogenous time series features alongside static features in a structured manner, while maintaining cohesive interpretation. Our approach leverages the insights of trajectory comprehension to introduce an encoding mechanism for exogenous time series, where they are decomposed into meaningful trends and properties, enabling the extraction of interpretable patterns. Through experiments on several synthetic datasets, we demonstrate that our approach remains predictive while preserving interpretability and robustness. This work represents a step towards developing robust, and generalized time series forecasting models. The code is available at https://github.com/jeremy-qin/TIMEVIEW