Pre-SMATS: A multi-task learning based prediction model for small multi-stage seasonal time series

Abstract

Learning on time series, especially on the small seasonal time series, has a wide range of practical applications. In this paper, to improve the learning effect on small seasonal time series, we extract the implicit information from the original series to enhance the data features and utilize stages to convey seasonality. Considering that many seasonal time series have implied stage characteristics, we propose Pre-SMATS, a multi-task learning based prediction model for small multi-stage seasonal time series. The model consists of three components: a feature extractor, a sub-task classifier and a main task predictor. The feature extractor is used to learn the feature vectors of the input data, the sub-task classifier is employed for stage classification and the main task predictor is applied for the final prediction. After extracting features of the input data, the extracted vectors are passed into the sub-task classifier and the main task predictor, respectively. Having learned the stage vector (i.e. the embedding of stage characteristics), the sub-task classifier outputs it to the main task, which concatenates the stage vector with the feature vector extracted by the feature extractor, to improve the prediction accuracy of the model. In order to verify our proposed model, Pre-SMATS, we conducted extensive experiments on two datasets with small time series, one is Chinese civil aviation passenger traffic (CCAPT) (2013–2019), a multi-stage seasonal time series and the other is IC board production furnace temperature curve (FTC), a general multi-stage time series. The experimental results show that our proposed model is superior to the baseline models in generalization capability as well as the accuracy of prediction on both two datasets. Specifically, compared with the best-performing baseline neural network model, the five error metrics (MSE, RMSE, MAE, MAPE and MSLE) of Pre-SMATS on CCAPT are reduced by 34%, 19%, 24%, 23% and 39% respectively, and on the FTC, they dropped by 42%, 25%, 16%, 18% and 43%.