Reveal the hidden layer via entity embedding in traffic prediction

Abstract

The neural network-based models have been widely used in traffic prediction. They have improved accuracy and efficiency in traffic flow, speed, passenger flow, and delay. Many variables are considered to predict traffic indicators and good techniques for choosing the most influenced variables to results have been developed. Since the neural network models treat independent variables as continuous variables, there are few studies on the use of categorical variables. In addition, the neural network has been criticized as the internal relationships of hidden layers are generally unknown. This paper investigates neural networks to predict the use of bike-sharing systems in Suzhou, China considering a large amount of categorical data. Two methods here, Entity embedding and one-hot encoding are applied. The comparison experiments verify that the entity embedding method is more efficient than one-hot encoding. Furthermore, the hidden layers are visually analyzed by t-SNE, and the relationships with time, weather, surroundings and other variables for the traffic volume at shared bike sites are discussed. The research results show that: 1. Entity embedding can effectively increase the continuity of categorical variables and therefore, improve the prediction efficiency for the neural network models. 2. The relationship between variables can be identified through visual analysis, and the trained embedding vectors can also be used to supervise clustering.