Variational embedding of a hidden Markov model to generate human activity sequences

Abstract

Although human trajectory data that are collected passively from location-based services (LBS) are regarded as a substitute for household travel surveys that entail a larger cost, the reality is that the data cannot be utilized directly for transportation planning and policy making without imputing missing qualitative information. Deep learning technologies have been widely used to infer the hidden features of passively collected mobile data. A deep neural network, however, is so deterministic that the probabilistic aspect of activity inference cannot be accommodated. In the present study, a stochastic approach (VAE-HMM) was devised to generate human activity chains by incorporating a variational autoencoder (VAE) with a hidden Markov model (HMM). Whereas an original HMM clusters data in the observational space, the proposed approach conducts clustering in a latent space with a smaller dimension. The VAE contributes by both reducing the input dimensionality and by sidestepping the overfit to sample data. The variational inference (VI) method was used to estimate the parameters of VAE-HMM within a Bayesian framework. Data drawn from spatio-temporal, demographic, socio-economic, and individual-specific sources were chosen as input variables to feed the model. The VAE-HMM can be trained in either a supervised or an unsupervised manner.