XLinear: Frequency-Enhanced MLP with CrossFilter for Robust Long-Range Forecasting

Researcher Xiang Ao has introduced XLinear, a new machine learning architecture designed to solve a persistent problem in time series forecasting: capturing long-range dependencies without sacrificing robustness. Traditional MLP (Multi-Layer Perceptron) models are known for their noise resistance but struggle with complex, long-term patterns, while Transformer models can capture dependencies but are often less robust and computationally heavy. XLinear bridges this gap by employing a dual-component approach, first decomposing a time series into trend and seasonal parts.

For the long-range trend component, XLinear uses a novel Enhanced Frequency Attention (EFA) module. This module operates in the frequency domain, a mathematical transformation of the data, which is more efficient for identifying repeating patterns over very long sequences. For the seasonal component, which contains shorter, noisier cycles, the model uses a proposed CrossFilter Block. This block is designed to filter out noise while preserving important signals, specifically avoiding the robustness issues common in attention mechanisms used by Transformers. Published and accepted at the 2025 IEEE AIAHPC conference, experimental results show XLinear achieves state-of-the-art performance, proving that a lightweight MLP-based structure can indeed be enhanced to master long-range forecasting tasks.