Reproduce “Few-Shot Forecasting of Time-Series with Heterogeneous Channels” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lukas Brinkmeyer, Rafael Rego Drumond, Johannes Burchert, Lars Schmidt-Thieme
TimeHetNet, a meta-learning model, facilitates few-shot forecasting of multivariate time series with heterogeneous channels, extending permutation-invariant methods to time-series data. It consistently outperforms traditional and meta-learning baselines across various forecasting horizons, while also introducing a public benchmark meta-dataset.
Key insights
- Explicitly incorporating temporal information through GRU and convolutional layers is crucial for improving short-to-medium horizon forecasts compared to static vector-based meta-learning methods.
- Meta-learning is essential for few-shot time-series forecasting, as training traditional models from scratch on minimal support data yields significantly inferior performance.
- Leveraging heterogeneous covariate channels is necessary for effective long-horizon time-series forecasting, particularly when historical target data is limited.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.