<ul><li>Proposed FaCTR, a lightweight spatiotemporal Transformer designed for time series forecasting with low per-timestep information density and complex dependencies.</li><li>FaCTR incorporates dynamic, symmetric cross-channel interactions through a learnable gating mechanism and encodes static and dynamic covariates for multivariate conditioning.</li><li>Despite its compact design with around 400K parameters, FaCTR achieves state-of-the-art performance on eleven public forecasting benchmarks for both short-term and long-term horizons.</li><li>The structured design of FaCTR enables interpretability through cross-channel influence scores and supports self-supervised pretraining, making it suitable for various time series tasks.</li></ul>

FaCTR: Factorized Channel-Temporal Representation Transformers for Efficient Time Series Forecasting

Discover more