PAMNET: Patch-Adaptive Mixing Network for Multivariate Time Series Forecasting

Multivariate Time Series Forecasting (MTSF) requires modeling temporal dynamics in a scalable and efficient way. Recent Transformer-based methods have achieved remarkable performance in MTSF, but they suffer from high computational cost due to the quadratic complexity. Although MLP-based models significantly reduce computational complexity, some of them explicitly model cross-channel interactions to enhance the model capacity, thereby introducing interference from irrelevant channels. Furthermore, shared decoders lead models to neglect important channel-specific patterns. In this paper, we propose a Patch-Adaptive Mixing Network called PAMNet to address these challenges. PAMNet first employs a Basis Augmented Patching (BAP) module to learn a compact low-rank basis and fuses it with local patch representations for each channel. Then, a Channel Adaptive Mixing Experts (CAME) module assigns channel-specific weights over a shared pool of experts via lightweight gating, enabling the model to capture heterogeneous temporal dependencies efficiently. This framework learns each channel’s unique temporal patterns without interference from irrelevant channels, and significantly reduces computational cost, ultimately improving forecasting performance. Extensive experiments on several real-world datasets demonstrate that PAMNet achieves state-of-the-art accuracy while maintaining low computational overhead.