[ICASSP 2023] Introducing a Multi-Resolution Sequence Aggregation and Model-Agnostic Framework for Time Series Forecasting

LG AI Research presented the MAMA(Multi-resolution sequence Aggregation and Model-Agnostic) framework at ICASSP 2023 to improve time series forecasting performance.

Existing multi-resolution approaches had limitations in that they either utilized only limited resolution information or lost the sequential order, which is a core characteristic of time series, during the process of merging information. To address this, MAMA proposes the following key modules.

• Multi-resolution input generation module: Captures fine-grained detail information by utilizing high-resolution data restored through a super-resolution upsampler.
• Model-agnostic forecasting module: Highly versatile, as it can be combined with various existing time series forecasting (TSF) algorithms.
• Sequential aggregation module: Preserves the sequential characteristics of time series when combining extracted information.
• Multi-resolution attention fusion module: Captures and combines the key temporal dynamics of each resolution signal through cross-attention.

In addition, this presentation also covers the latest Causal research trends, including interpretable multi-scale neural networks for Granger causal discovery and research on sound event detection through causal intervention.