A Novel Feature Emphasis Mechanism on Traffic Flow Prediction

Traffic flow prediction is a challenging application in temporal prediction, especially in a small amount of dataset. Recently, most of the studies are based on a large data set, which will lead the huge waste of resources and time. In the case of small data sets, some state-of-the-art models cannot achieve the desired results due to the large deficiency of features. In this work, the main idea is to use our proposed Feature Emphasis Mechanism (FEM) to augment, re-extract and emphasis the small amount features by using Linear-Emphasis (LE) and Conv-Emphasis (CE). To capture and emphasis features more effectively, we apply Bi-training to capture the parameters in model. Extensive experiments demonstrate that Feature Emphasis Mechanism can be applied to classical models, and improves the prediction accuracy in small dataset background.