Reproduce “MoTM: Towards a Foundation Model for Time Series Imputation based on Continuous Modeling” on OpenResearch

About this paper

Etienne Le Naour, Tahar Nabil, Ghislain Agoua

MoTM introduces a foundation model for time series imputation, leveraging a mixture of continuous Implicit Neural Representations to address irregular sampling and out-of-domain generalization. The approach achieves competitive to superior imputation performance against supervised methods and significantly outperforms other zero-shot foundation models across various real-world datasets, while maintaining inference efficiency.

Key insights

• Combining a mixture of independently trained continuous-time models with an adaptive orchestrator enables robust out-of-domain generalization for time series imputation.
• Implicit Neural Representations naturally accommodate heterogeneous sampling rates and irregular time grids, making the framework adaptable to diverse real-world time series.
• The meta-learning nature of TimeFlow and the closed-form solution of ridge regression allow for efficient adaptation and combination of expert predictions at inference time.

Summary by alphaXiv.