Exploring variabilities of multi-week activity-travel patterns: a deep embedded clustering approach

A good understanding of activity-travel pattern (ATP) variabilities in transit networks can help government or transit operators improve the accuracy of travel demand forecasts and adjust transport supply. Previous studies on ATP variabilities are often limited to a short period and cannot comprehensively reflect the interpersonal and intrapersonal variabilities. In addition, the traditional clustering methods lack sufficient learning ability for high-dimensional feature space, and clustering variables are simple aggregated indicators. In this study, we propose an ATP inference algorithm that reconstructs multi-week individuals’ discrete trips into an ATP, and the ATP is modeled as stochastic process. Various indicators regarding the standard deviation (SD) of travel time, the SD of number of trips, and the entropy of ATPs are applied, and the entropy rate of ATPs is explicitly considered to take account of the order of activity/travel choices. The intrapersonal variability of ATPs is obtained through these indicators. The interpersonal variability of ATPs is investigated by dividing people into groups based on their ATPs through a deep embedded clustering approach, which refers to a deep neural network formed by integrating the stacked denoising autoencoder with the basic clustering algorithm. The proposed deep embedded clustering is tested using massive smart card data collected from the metro system in Nanjing, China. Comparative experiments validated that the variabilities of multi-week ATPs investigated by the deep embedded clustering approach can help realize a more accurate ATP prediction.