Exploring the Role of Token in Transformer-based Time Series Forecasting (2404.10337v3)

Abstract: Transformer-based methods are a mainstream approach for solving time series forecasting (TSF). These methods use temporal or variable tokens from observable data to make predictions. However, most focus on optimizing the model structure, with few studies paying attention to the role of tokens for predictions. The role is crucial since a model that distinguishes useful tokens from useless ones will predict more effectively. In this paper, we explore this issue. Through theoretical analyses, we find that the gradients mainly depend on tokens that contribute to the predicted series, called positive tokens. Based on this finding, we explore what helps models select these positive tokens. Through a series of experiments, we obtain three observations: i) positional encoding (PE) helps the model identify positive tokens; ii) as the network depth increases, the PE information gradually weakens, affecting the model's ability to identify positive tokens in deeper layers; iii) both enhancing PE in the deeper layers and using semantic-based PE can improve the model's ability to identify positive tokens, thus boosting performance. Inspired by these findings, we design temporal positional encoding (T-PE) for temporal tokens and variable positional encoding (V-PE) for variable tokens. To utilize T-PE and V-PE, we propose T2B-PE, a Transformer-based dual-branch framework. Extensive experiments demonstrate that T2B-PE has superior robustness and effectiveness.