Predicting Human Mobility on Large-scale Mobility Traces

Abstract

Human mobility prediction has attracted a lot of attention because it plays the key point for the success in a variety of applications ranging from location-based recommendation systems to epidemiology. Therefore, the purpose of this thesis is to answer two main questions in human mobility prediction: where a person is most likely to visit and whom a person is most likely to meet. Accordingly, two prediction models which estimate the most probable future locations and encounters are proposed.

In the first model, we aim at estimating the next location of a person-of-interest even when the recent information about the position of that person is unknown. Motivated by the fact that the behavior of an individual is greatly related to other people, a two-phase framework is proposed, which first finds persons who have highly correlated movements with a person of interest, then leverages the position information of selected persons to predict the person-of-interest's location. For the first phase, we propose two methods: community interaction similarity-based (CISB) and behavioral similarity-based (BSB). The CISB method finds persons who have similar encounters with other members in the entire community. In the BSB method, members are selected if they show similar behavioral patterns with a given person, even though there are no direct encounters or evident co-locations between them. For the second phase, a neural network is considered in order to develop the prediction model based on the selected members.

The purpose of the second model is to design a low cost, high accurate human encounter prediction framework that can be applied to large-scale networks. Taking inspiration from the advantages of the distributed system (e.g., low cost and ease of scaling up) and the temporal dependency of human mobility, we propose the distributed human encounter prediction (DHEP) model, which uses the mobility history of only the person of interest to estimate future encounters of that person. The DHEP model based on a recurrent neural network is constructed, in which recent encounter information is captured and used to make future contact predictions. In addition, for devices with constrained computation capability, we design a feed-forward neural network-based DHEP model which contains a much smaller number of model parameters. Also, an embedding model that learns the low-dimensional representation of a person's location is proposed in order to accelerate the training of the prediction model.

Extensive experiments have been conducted to evaluate the performance of the proposed prediction models with a variety of parameters on different large mobility traces (e.g., MIT, Dartmouth, and UB datasets). The evaluation results show that the designed frameworks achieve higher performance in terms of predictive accuracy than existing studies. Specifically, the human location prediction model under the BSB method outperforms other selection methods on considered datasets. Moreover, the decentralized encounter prediction model with low overhead and high accuracy can protect data privacy and can be applied to large-scale networks.