Title: SEMixer: Semantics Enhanced MLP-Mixer for Multiscale Mixing and Long-term Time Series Forecasting

Abstract:

Accurately capturing multiscale patterns is essential for effective long-term time series forecasting (TSF). Nevertheless, the presence of noise and redundancy within time series data, coupled with semantic discrepancies between non-adjacent scales, creates significant hurdles for efficiently aligning and integrating multi-scale temporal dependencies. To overcome these challenges, we introduce SEMixer, a lightweight model specifically engineered for long-term TSF tasks.

SEMixer incorporates two primary innovations: a Random Attention Mechanism (RAM) and a Multiscale Progressive Mixing Chain (MPMC). During the training phase, RAM identifies diverse interactions among time patches, which are then aggregated through dropout ensemble techniques at inference. This process enriches patch-level semantics, allowing the underlying MLP-Mixer architecture to more effectively model multi-scale dependencies. Additionally, MPMC efficiently stacks RAM and the MLP-Mixer to achieve superior temporal mixing while maintaining low memory consumption. This design mitigates semantic gaps across various scales, thereby improving both multiscale modeling capabilities and overall forecasting accuracy.

We demonstrate SEMixer’s efficacy across ten public datasets. Furthermore, the model secured third place in the \textit{2025 CCF AlOps Challenge}, leveraging a substantial dataset of 21GB of real-world wireless network data. The source code is publicly accessible at https://github.com/Meteor-Stars/SEMixer.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC