A new paradigm of time series forecasting with attributes learning

The objective loss function pertains to the training approach and the effectiveness of the model. However, the widely used distances-based loss functions, which do not consider the dynamic temporal pattern of time series, resulting in its inability to capture the trends and seasons of the sequence well. Additionally, in practical applications, time series are often attacked by noise, resulting in various distortion phenomena, so how to combat the problem of data distortion is a challenge in loss function design. Aiming at the above problems, a loss function framework was designed based on time series attribute (including trend and season) learning, which consists of trend direction guidance, seasonal change indication and point-wise representation terms. Based on the loss function framework, a novel loss function Attri-Loss (loss function based on attribute learning), was proposed for time series forecasting model, that not only considers the distortions in all aspects but also allows models to capture the attributes of time series. Attri-Loss is a plug-and-play loss function that can be directly applied to almost arbitrary time series forecasting network, which is Model-Agnostic. We evaluated the effectiveness of Attri-Loss by conducting extensive experiments from naive models to state-of-the-art models. The experimental results show that Attri-Loss not only reduces the prediction error but also improves the shape similarity. Compared with state-of-the-art loss functions for 72-hour prediction, it achieved the best predictive performance, which the mean absolute error decreased by about 13%, and the symmetric mean absolute percentage error decreased by about 8%. The code and models are available at GitHub https://github.com/liaohaibing/attribute-learning-loss.