LG AI Research 334

LG AI Research participated in ICASSP 2023, held in Rhodes, Greece, and introduced the MAMA (Multi-resolution sequence Aggregation and Model-Agnostic) framework for time series forecasting.

Existing multi-resolution approaches had limitations in that they either utilized only limited resolution information or lost the sequential characteristics essential to time series data during the information merging process. To overcome this, the MAMA framework proposes the following key modules.

• Multi-resolution input generation: Generates high-resolution signals through a super-resolution upsampler to leverage fine-grained details.
• Model-agnostic forecasting: Improves performance by adding multi-resolution information while still using existing time series forecasting (TSF) algorithms as they are.
• Sequential aggregation: Preserves the sequential characteristics of time series when combining the extracted multi-resolution information.
• Multi-resolution attention fusion: Captures and combines the key temporal dynamics of each resolution signal through cross-attention.