Recombination-based two-stage out-of-distribution detection method for traffic flow pattern analysis

There is numerous research on traffic flow pattern analysis from the temporal or spatial perspective by leveraging signals from heterogeneous information sources such as road sensors, traffic cameras, and so on. However, most of the existing studies focus on the patterns that appeared in the historical traffic data and rarely consider the potential traffic flow pattern (or distribution) shifts where the distribution of the newly arrived data is different from that of the original data. The shift of the data pattern can be caused by various reasons such as fundamental changes in travel behaviour (e.g., COVID-19), road network changes (e.g., adding a new junction), or other infrastructure changes (e.g., relocation of a school). In this chapter, a novel supervised classification method with the ability to detect potential traffic flow pattern shifts is proposed based on a neural network architecture. The detection of pattern shifts is achieved by quantitative uncertainty modelling using a newly developed recombination-based machine learning method. Therefore, a new sample from the original data distribution will yield a prediction with a low uncertainty score while a sample from the shifted distribution gives a high uncertainty score. From a statistical perspective, the proposed method measures the relative distance between a set of recombined sample pairs generated by the new sample and the original dataset to discover the statistical features for uncertainty computation of the predicted outcomes. A case study using real data from Cambridge, UK, shows that the proposed approach achieves accuracy of 97.79% and 92% on the prediction for the test dataset and OOD dataset, respectively, and is capable of detecting shifts in overall traffic flow patterns over the COVID-19 pandemic period.