Saxformer: A Time Series Forecasting Framework with Local Interpretability
Song Ying, Danjing Li, Kai Sun, Yin Zheng · 2024
In the era of information, the importance of time series data has become increasingly prominent. However, existing forecasting models often face the challenge of insufficient interpretability, leading to potential risks in practical applications. In this paper, we propose Saxformer, a novel forecasting model which combines the Symbolic Aggregate Approximation symbolization method with Transformer architecture. By visualizing its output matrices, the local interpretability of the model is intuitively presented. In addition, a 1-D convolution-and-deconvolution structure is incorporated with the model to better capture localized information of time series data. Experiment results show that Saxformer outperforms the existing mainstream models with multiple evaluation metrics, particularly excelling in the forecasting of low-dimensional time series data.