Individualized Passenger Travel Pattern Multi-Clustering based on Graph Regularized Tensor Latent Dirichlet Allocation

Award:

This paper is awarded with Best Student Paper Finalist Award in INFORMS 2020 Data Mining Section.

After the previous two works in station-wise traffic flow prediction, we realized this mass traffic analysis at a macro level neglects the research value and abundant information of individual passenger travel data. Individualized travel pattern (passenger $u$ travels from origin $o$ to destination $d$ at time $t$) is believed to have higher research value.

Challenge

But this task is rather challenging since it is high-dimensional multi-mode Spatiotemporal big data with more 7 million passengers; Also there is multi-clustering structure along each dimension of $o, d, t$; passenger behaviors are also affected by the external environment, such as the locations and surroundings of stations.

Methodology

So we proposed a novel Graph-Regularized Tensor LDA model: firstly it represents each trip from one passenger as a 3-dimensional word $\boldsymbol{w} = (w^O, w^D, w^T)$; A passenger with several trips is perceived as a 3-dimensional document $\boldsymbol{\mathcal{W}}^{O \times D \times T}$; Generative processes in the passenger-level and trip-level will be defined along each dimension and the latent topic will be also formulated as a tensor $\boldsymbol{z} = (z^O, z^D, z^T)$; Same as last work we also observed that passengers will have similar patterns both in geographically close stations or functionally similar stations. We further propose to incorporate the graph regularizations into the tensor LDA generative process for origin and destination. We also propose the tensorised variational expectation-maximization (EM) algorithm to estimate parameters.

Results

Eventually, the topics along each dimension of $o, d, t$ are much more interpretable and meaningful than benchmark methods.