Time-Series Forecasting With Semantic-Aware Trend Loss Function

Time-series forecasting (TSF) is becoming ubiquitous in numerous real-world applications. The loss function is a vital component of time-series forecasting models based on deep learning. The widely usedLp-norm distances-based loss functions, which are vulnerable and biased to not consider dynamic temporal patterns and label autocorrelation, resulting in its inability to capture the semantics or shape of the sequence well. In practice, the sequence data will have various distortions due to the influence of environmental factors, and how to overcome the data distortion variations is an important challenge for loss function design in TSF. Aiming at the above challenges, a loss function framework based on semantic-aware learning was design, which consists of global trend direction guidance and point-wise local trend terms. Based on the loss function framework, we propose a novel loss function, called Tre-Loss (semantic-aware trend learning based loss function), that not only considers the distortions in all aspects but also allows models to capture the semantics of time-series. Tre-Loss is a plug-and-play loss function that can be directly applied to almost arbitrary TSF network. Evaluate the effectiveness of Tre-Loss by conducting extensive experiments from naive models to state-of-the-art models. The experiment results indicate that the models trained with Tre-Loss outperforms those trained with other training metrics. The code and models will be available at GitHub.