Search papers, labs, and topics across Lattice.
The paper introduces Hierarchical Patching Mixer (HPMixer), a novel architecture for long-term multivariate time series forecasting that explicitly models periodic patterns and residual dynamics in a decoupled manner. HPMixer uses a learnable cycle module with a nonlinear channel-wise MLP for periodicity and a Learnable Stationary Wavelet Transform (LSWT) followed by a channel-mixing encoder and hierarchical patching for residuals. Experiments on standard benchmarks demonstrate that HPMixer achieves competitive or state-of-the-art performance.
By decoupling periodicity and residuals with hierarchical patching, HPMixer achieves state-of-the-art multivariate time series forecasting without complex attention mechanisms.
In long-term multivariate time series forecasting, effectively capturing both periodic patterns and residual dynamics is essential. To address this within standard deep learning benchmark settings, we propose the Hierarchical Patching Mixer (HPMixer), which models periodicity and residuals in a decoupled yet complementary manner. The periodic component utilizes a learnable cycle module [7] enhanced with a nonlinear channel-wise MLP for greater expressiveness. The residual component is processed through a Learnable Stationary Wavelet Transform (LSWT) to extract stable, shift-invariant frequency-domain representations. Subsequently, a channel-mixing encoder models explicit inter-channel dependencies, while a two-level non-overlapping hierarchical patching mechanism captures coarse- and fine-scale residual variations. By integrating decoupled periodicity modeling with structured, multi-scale residual learning, HPMixer provides an effective framework. Extensive experiments on standard multivariate benchmarks demonstrate that HPMixer achieves competitive or state-of-the-art performance compared to recent baselines.
