HyperAI

Past Decomposable Mixing

Past Decomposable Mixing (PDM) is a theory proposed by Shiyu Wang, Haixu Wu and others. This theory was published in the paper "TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting"It was first proposed in .

PDM is a theoretical concept for time series forecasting and one of the core components of the TimeMixer model. The design of PDM is based on the observation that time series present different patterns at different sampling scales, and it extracts details and macro trends from historical information by decomposing and mixing seasonal and trend components at different scales. Specifically, the PDM module decomposes the time series into seasonal and trend components, and mixes these components in the fine-to-coarse and coarse-to-fine directions, respectively, aggregating micro seasonal and macro trend information in turn.

The main contributions and features of PDM include:

  1. Multi-scale mixed perspective: PDM exploits the variations of time series data at different time scales to handle complex temporal variations in time series forecasting by decoupling variations and complementary forecasting capabilities from multi-scale series.
  2. Decoupling past information: The PDM module can decouple the past information of the time series and extract seasonal and trend characteristics, so as to better understand and predict the future behavior of the time series.
  3. Efficient running time: Since the TimeMixer model is based on a full MLP (multi-layer perceptron) architecture, PDM exhibits good efficiency at runtime and is suitable for real-time or near real-time prediction scenarios.

PDM combined with the Future-Multipredictor-Mixing (FMM) module enables TimeMixer to achieve consistent state-of-the-art performance in both long-term and short-term prediction tasks with good runtime efficiency.