Understanding Origin-Destination Ride Demand with Interpretable and Scalable Nonnegative Tensor Decomposition.

This paper focuses on the estimation and compression of ride demand from origin-destination (OD) trip event data. By representing the OD event data as a three-way tensor (origin, destination, and time), we model the data as a Poisson process with an intensity tensor that can be decomposed according to a Tucker decomposition. We establish and justify a specific form of nonnegative Tucker-like tensor decomposition that represents OD demand via K latent origin spatial factors and K latent destination spatial factors. We then provide a computational and memory efficient algorithm for performing this decomposition and demonstrate its use for real-time compression and estimation of OD ride demand. Two case studies based on New York City (NYC) taxi and Washington DC (DC) taxi were implemented. Results from the case studies demonstrate the applicability of the proposed method in data compression and short-term forecast for ride demand. Furthermore, we found that the learned latent spatial factors are interpretable and localized to specific areas for both NYC and DC cases. Hence, this method can be used to understand OD trip data through latent spatial factors and be used to identify spatio-temporal patterns for OD trip and travel demand generation mechanism in general. Funding: This work was supported by the U.S. Department of Transportation [UTC/NCST] and the U.S. National Science Foundation [Grant DMS 1712996]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/trsc.2022.0101. [ABSTRACT FROM AUTHOR]