Granularity Fusion Transformer: Learning multi-granularity patterns for time-series forecasting

Jinwoo Park, Hyeongwon Kang, Seung Hun Han, Pilsung Kang

Abstract: Time series data consist of continuous observations collected over time. Theoretically, time series data form a continuous trajectory, but in practice, the actual data used in the real-world are discrete time series obtained by sampling from a continuous trajectory. By considering the sampling interval, it is possible to collect data with different levels of information, referred to as granularity in time series data. Most existing studies assume single-granularity, which leads to failure in capturing variability occurring at different levels of granularity. To address these issues, we propose the Granularity Fusion Transformer (GFT), which addresses characteristics at different granularity levels. GFT estimates multi-granularity data by simultaneously considering the information of period and phase at the single-granularity level. To merge granularities that are segregated into different levels, we propose a patch-wise cross-attention based Granularity Fusion Encoder. Extensive experiments on six datasets demonstrate that the proposed method outperformed benchmark models by reducing MSE by 35.27% and MAE by 22.80%, thereby achieving more accurate predictions closer to the actual values. These results highlight the usefulness of multi-granularity in time series forecasting tasks.