NPMixer: Hierarchical Neighboring Patch Mixing for Time Series Forecasting

2026-05-08 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

NPMixer, or Neighboring Patching Mixer, is a new hierarchical architecture for multivariate time series forecasting, introduced on May 8, 2026. It addresses the challenges of local temporal dynamics and global dependencies by incorporating a Learnable Stationary Wavelet Transform, which adaptively decomposes signals into trend and detail components. The framework features a Neighboring Mixer Block that uses hierarchical MLP layers on non-overlapping patches to capture local temporal dynamics and expand the receptive field for multi-scale dependencies. Additionally, a Channel-Mixing Encoder processes high-frequency components to learn channel correlations while maintaining global trend stability. Extensive experiments across seven benchmark datasets show NPMixer outperforms existing models, achieving superior Mean Squared Error (MSE) performance in 20 out of 28 (71.4%) experimental setups.

Key takeaway

For AI Engineers and Research Scientists developing multivariate time series forecasting models, NPMixer offers a robust architecture that significantly improves accuracy. You should consider integrating hierarchical patching with adaptive wavelet transforms and channel-mixing encoders into your next-generation forecasting solutions, especially when dealing with complex local dynamics and global dependencies. This approach can lead to more stable and precise predictions across diverse datasets.

Key insights

NPMixer uses hierarchical patching and adaptive wavelet transforms for superior multivariate time series forecasting.

Principles

• Decompose signals into trend and detail components.
• Capture local dynamics with hierarchical MLP layers.
• Preserve global trend stability while learning channel correlations.

Method

NPMixer employs a Learnable Stationary Wavelet Transform for signal decomposition, followed by a Neighboring Mixer Block with hierarchical MLPs on patches, and a Channel-Mixing Encoder for high-frequency components.

In practice

• Apply adaptive wavelet transforms for data-dependent signal decomposition.
• Utilize MLP-based patching for local temporal pattern extraction.
• Integrate channel mixing for high-frequency components.

Topics

• NPMixer
• Time Series Forecasting
• Multivariate Time Series
• Hierarchical Patch Mixing
• Learnable Wavelet Transform

Code references

• Meteor-Stars/SEMixer
• Sachithx/EntroPE

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Conformal Prediction with Time-Series Data via Sequential Conformalized Density Regions — SCDR offers doubly robust, asymptotically valid conformal prediction for time-series, generating informative, potentially disconnected prediction sets.
• RegMix-D: Dynamic Data Mixing via Proxy Training Trajectories — Leveraging full loss trajectories from proxy runs enables dynamic data mixture optimization for LLM pretraining.
• Multiple cyclicity and Wavelet Decomposition with Channel Correlation for Long-term Time Series Forecasting — McWC improves long-term time series forecasting by explicitly modeling cyclicity, trend, and inter-channel correlations with wavelet decomposition.
• A Streaming Sparse Cholesky Method for Derivative-Informed Gaussian Process Surrogates Within Digital Twin Applications — Derivative-informed sparse Gaussian Processes enable accurate, dynamically updated digital twin surrogates for real-time system prognostics.
• Zero Touch Predictive Orchestration: Automating Time-Series Models for the Cloud-Edge Continuum — Automated data-mixing with a foundational dataset mitigates cold start for time-series forecasting in volatile Cloud-Edge Continuum environments.
• Towards a Unified Generative Model for Scarce Time Series with Domain Experts — TimeMoDE unifies Diffusion Transformers and Mixture-of-Experts for robust time series generation in data-scarce settings.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.