FSMLP: Modelling Channel Dependencies With Simplex Theory Based Multi-Layer Perceptions In Frequency Domain

ICLR 2026 Conference Submission16825 Authors

Keywords: time series forecasting

Abstract:

Time series forecasting (TSF) plays a crucial role in various domains. While effective for temporal modeling, channel-wise MLPs suffer from overfitting in inter-channel dependency learning. In this paper, we analyze this via Rademacher complexity theory, identifying extreme values as key overfitting catalysts.
To mitigate this issue, we propose to constrains weights to a standard simplex (Simplex-MLP), enforcing simpler patterns and reducing extreme value overfitting. Theoretically, we demonstrate that Simplex-MLP exhibits reduced susceptibility to overfitting on extreme values and demonstrates enhanced generalization capabilities. Based on the Simplex-MLP layer, we propose a novel F requency S implex MLP (FSMLP) framework for time series forecasting, comprising of two kinds of modules: Simplex Channel-Wise MLP (SCWM) and FrequencyTemporal MLP (FTM). Experiments on seven benchmarks confirm FSMLP's accuracy/efficiency improvements and superior scalability. Additionally, simplex-MLP also enhances existing channel-wise MLP methods, reducing their overfitting and boosting performance. Code is available.

Primary Area: learning on time series and dynamical systems

Submission Number: 16825