Application of the Affinity Propagation Clustering Technique to obtain traffic accident clusters at macro, meso, and micro levels

Accident grouping is a crucial step in identifying accident-prone locations. Among the different accident grouping modes, clustering methods present excellent performance for discovering different distributions of accidents in space. This work introduces the Affinity Propagation Clustering (APC) approach for grouping traffic accidents based on criteria of similarity and dissimilarity between distributions of data points in space. The APC provides more realistic representations of the distribution of events from similarity matrices between instances. The results showed that when representative data samples obtain, the preference parameter of similarity provides the necessary performance to calibrate the model and generate clusters according to the desired characteristics. In addition, the study demonstrates that the preference parameter as a continuous parameter facilitates the calibration and control of the model's convergence, allowing the discovery of clustering patterns with less effort and greater control of the results