SEMixer: Semantics Enhanced MLP-Mixer for Multiscale Mixing and Long-term Time Series Forecasting (In ACM Web Conference 2026, WWW 2026)

Datasets

You can download the public datasets used in our paper from https://drive.google.com/drive/folders/1PPLsAoDbv4WcoXDp-mm4LFxoKwewnKxX. The downloaded folders e.g., "ETTh1.csv", should be placed at the "dataset" folder. These datasets are extensively used for evaluating performance of various time series forecasting methods.

Overview

• We propose an end-to-end lightweight multiscale model, SEMixer, for long-term forecasting. Due to meticulous model design, SEMixer can handle longer input sequences and achieve better performance from longer inputs.

• We propose the Random Attention Mechanism (RAM), which learns diverse interactions through random sampling during training and integrates them for enhancing the semantics of time patches through dropout ensemble, resulting in higher efficiency and better forecasting performance.

• We propose the Multiscale Progressive Mixing Chain (MPMC) to progressively stack the RAM and MLP-Mixer backbone as the TS scale level increases, and make them only work in the pairwise concatenation of adjacent scales. This not only facilitates better forecasting performance due to considering the semantic gaps among different scales, but also results in low memory usage because MLP-Mixer does not need to process all scale inputs at once. We also observe that MPMC helps to resist noise.