The paper proposes a novel Channel Clustering Module (CCM) that addresses the limitations of existing channel strategies in time series forecasting. Key highlights:
Motivation: Observation that the model's reliance on channel identity information is anti-correlated with the similarity between channels. This suggests the potential to replace channel identity with cluster identity.
Channel Clustering: CCM learns to dynamically group channels based on their intrinsic similarities, creating a clustering membership matrix. It also learns expressive prototype embeddings for each cluster using a cross-attention mechanism.
Cluster-aware Feed Forward: Instead of using individual or shared Feed Forward layers, CCM assigns a separate Feed Forward layer to each cluster to capture the underlying shared time series patterns within the clusters.
Zero-shot Forecasting: The learned prototypes enable zero-shot forecasting on unseen samples by grouping them into appropriate clusters.
Experiments: CCM consistently improves the performance of four state-of-the-art time series models across long-term and short-term forecasting benchmarks, as well as enhances their zero-shot forecasting capabilities in cross-domain and cross-granularity scenarios.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Jialin Chen,... at arxiv.org 04-03-2024
https://arxiv.org/pdf/2404.01340.pdfDeeper Inquiries