What is it about?
TSDE (Time Series Diffusion Embedding) is a novel Time Series Representation model, leveraging a denoising diffusion process at its core along with Transformer encoders. The model is trained in a self-supervised learning fashion with IIF masking strategy encapsulating imputation, interpolation and forecasting tasks.
Featured Image
Photo by Hunter Harritt on Unsplash
Why is it important?
This work explores how diffusion models could be used for time series representation learning, and tackles the challenges of handling multivariate time series (inter-features dependency) by employing two transformer encoders, temporal and spacial, to process both sequences of timely recorded data and different features' interactions. The model could be applied to a range of different time series tasks including imputation, interpolation, forecasting, anomaly detection, classification and clustering.
Perspectives
Writing this paper allowed me to explore new frontiers in time series modeling through the integration of diffusion processes and self-supervised learning. I'm proud of the innovative approach we developed with TSDE and hope it opens doors for further research in the field. Collaborating with such a talented team made this journey particularly rewarding.
Lele Cao
Microsoft Corp
Read the Original
This page is a summary of: Self-Supervised Learning of Time Series Representation via Diffusion Process and Imputation-Interpolation-Forecasting Mask, August 2024, ACM (Association for Computing Machinery),
DOI: 10.1145/3637528.3671673.
You can read the full text:
Contributors
The following have contributed to this page
