SSTGCN: A Deep Learning Framework for Road Intersection Similarity Learning

Accurate and real-time traffic road intersection feature extraction and similarity learning play an important role in the urban traffic control system while traditional signal timing requires a lot of manpower and time cost. We consider measuring the feature similarity between different intersections, to facilitate the traffic signal optimization strategy transfer of similar intersections. However, existing road intersection similarity learning methods are often distance-based measurement schemes, which are difficult to comprehensively measure spatio-temporal multivariate data along with a large amount of computation. Therefore, we propose a Siamese-Spatio-Temporal Graph Convolutional Networks (SSTGCN) with a heterogeneous multi-granularity aggregation strategy to capture the underlying spatial correlations and temporal dependencies among multi-hop intersections. The experimental results show that the proposed algorithm can accurately predict the similarity of two intersections with 47.61% lower RMSE and 12.04% higher accuracy compared with baseline. Furthermore, it is suitable for transferring the optimized intersection traffic strategy with SSTGCN.