Mixing It Up: Exploring Mixer Networks for Irregular Multivariate Time Series Forecasting

TL;DR

This paper introduces IMTS-Mixer, a novel neural network architecture designed for irregular multivariate time series forecasting, combining innovative encoding and decoding components to improve accuracy and efficiency.

Contribution

It adapts Mixer models to irregular time series by proposing ISCAM and ConTP components, achieving state-of-the-art results with fewer parameters.

Findings

State-of-the-art forecasting accuracy

Faster inference times

Fewer parameters than baselines

Abstract

Forecasting irregularly sampled multivariate time series with missing values (IMTS) is a fundamental challenge in domains such as healthcare, climate science, and biology. While recent advances in vision and time series forecasting have shown that lightweight MLP-based architectures (e.g., MLP-Mixer, TSMixer) can rival attention-based models in both accuracy and efficiency, their applicability to irregular and sparse time series remains unexplored. In this paper, we propose IMTS-Mixer, a novel architecture that adapts the principles of Mixer models to the IMTS setting. IMTS-Mixer introduces two key components: (1) ISCAM, a channel-wise encoder that transforms irregular observations into fixed-size vectors using simple MLPs, and (2) ConTP, a continuous time decoder that supports forecasting at arbitrary time points. In our experiments on established benchmark datasets we show that our model achieves state-of-the- art performance in both forecasting accuracy and inference time, while using fewer parameters compared to baselines.