Traffic prediction with embeddings

Forecasting traffic flow is a critical task in transportation and community transport. It involves predicting traffic speed in a road network using historical data. To do this, sensors measure and record the traffic speed on the chosen roads. However, this task is challenging because the sensors’ dependencies cannot be explained only by their relative positions in the Euclidean space. Moreover, traffic speeds highly depend on the day and time. In our experiments, we employed attributed and structural graph embeddings to capture node relationships. One of the advantages of using this form of graph representation is that the embedding can be trained in an
unsupervised manner, even when node attributes are assigned.