Reproduce “Self-supervised learning for fast and scalable time series hyper-parameter tuning” on OpenResearch

About this paper

Peiyi Zhang, Xiaodong Jiang, Ginger M Holt, Nikolay Pavlovich Laptev, Caner Komurlu, Peng Gao, Yang Yu

Researchers from Purdue University and Facebook Inc. developed self-supervised learning frameworks for time series model selection (SSL-MS) and hyper-parameter tuning (SSL-HPT). These frameworks predict optimal models and hyper-parameters directly from time series features, achieving 6-20 times faster hyper-parameter tuning than search-based methods while maintaining comparable or superior forecasting accuracy.

Key insights

• Time series features can effectively capture underlying characteristics sufficient to predict suitable forecasting models and their optimal hyper-parameters.
• Transforming iterative, search-based hyper-parameter tuning into a constant-time inference problem using self-supervised learning significantly reduces computational cost without degrading accuracy.
• The computational cost for training the self-supervised learner is empirically bounded, suggesting that a relatively small training set can generalize effectively to very large datasets, ensuring scalability.

Summary by alphaXiv.