This article extends the experimental benchmark from Wu et al. (2023) to evaluate time series modeling tasks like forecasting, classification, imputation, and anomaly detection under low-cost conditions. The benchmark contrasts with TimesNet by focusing on limited compute and supervision resources, simulating real-world scenarios. Classification tasks utilize unsupervised representation learning, and performance is evaluated based on zero-shot settings and linear probing methods. The datasets employed for benchmarking include standard datasets for forecasting and newly curated datasets for classification and anomaly detection from UCR archives.
Our benchmark comprises of 5 major time series modeling tasks of significant practical value, namely long- and shorthorizon forecasting, imputation, classification, and anomaly detection.
We exclusively consider scenarios characterized by limited compute and supervision resources, mimicking practical situations where training deep neural networks is infeasible.
Collection
[
|
...
]