Abstract:
Traffic prediction is an essential and challenging task for traffic management and commercial purposes. Machine learning methods for traffic prediction usually treat traffic conditions as time-series due to obvious temporal patterns. Recently, spatial relationships among roads in a road network have also been used to improve traffic prediction. This study proposes a novel method to predict traffic conditions such as speed using a graph convolutional neural network with a spectral adjacency matrix (GCN-Spectral). Unlike a spatial adjacency matrix representing physical connections between road segments, a spectral matrix represents the correlation between road segments regarding traffic conditions. The GCN-Spectral model is evaluated by comparing with a multi-layer perceptron model (MLP), as a non-spatial model, and a graph convolutional neural network with a spatial adjacency matrix (GCN-Spatial). The prediction results were analyzed with the robustness characteristics of the road segment in various dimensions. For example, the road length, time of the day, and day of the week. The error of results analysis aimed to explain model limitations and strong points. The data used in this study are GPS probe data collected from taxis in Bangkok. Empirical results show that the GCN-Spectral with a combination matrix model mostly outperforms GCN-Spatial models in the Bangkok dataset.
However, MLP performs the best in most cases in speed prediction tasks. The MLP works well every day of the week and time of day. In contrast, the GCN works well in late morning, evening, and on a weekday. The number of lanes in a road segment does not correlate with prediction error. And the road segment length has a weak correlation with the prediction error on GCN-Spectral with LSTM layers and GCN-Spectral with combination matrix. The travel time spent in the road segment is calculated using speed prediction, and relative to road segment length. The more extended the road segment is, the higher the error on travel time. The result found that the lowest error is from GCN-Spatial model.
