Latent Pattern Extraction Across Multi-Dataset Shared Mobility Data: Correspondence Finding Using Multi-Tensor Decomposition

Shared mobility services (such as electric scooters and ride shares) and the data they generate give us a unique opportunity to understand human mobility by combining trips with contextual data. Between-datasest correspondences shed light on the connections between groups of elements (such as people, places, or weather conditions) present in intelligent transportation systems (ITS) and related data. The prevalence of scooters and scooter datasets in multiple cities combined with their importance in today's urban life make scooter trips one such ITS example. One way between-dataset understanding can be achieved is with tensors, or multi-dimensional arrays, which provide a framework for simultaneously examining multiple datasets, such as subway traffic, scooter trips, demography, or weather data. Starting with a multi-tensor framework, we propose to use multi-tensor decomposition to extract latent patterns. These latent patterns are in turn used to cluster similar elements within data types and find correspondences between clusters in different datasets. We test Multi-Dataset Correspondence Finding (MDCF) on both synthetic data and empirical datasets of dockless vehicle trips in Minneapolis, Minnesota, and Louisville, Kentucky. On the empirical data, MDCF reveals connections between neighborhoods and their times of peak activity across both cities, both supporting prior research and offering new insights.