Effective Learning from Mobile Data for Human Behavior and Urban Dynamics Sensing and Prediction

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Effective Learning from Mobile Data for Human Behavior and
Urban Dynamics Sensing and Prediction"

By

Mr. Jiangchuan ZHENG


Abstract

The phenomenal growth of sensor-equipped smartphones and GPS-equipped vehicles 
have produced an unprecedented wealth of digital information, such as humans' 
phone call records, mobility traces, Bluetooth proximity readings, as well as 
city taxi trajectories. Learning the hidden patterns from such mobile data can 
help understand the contexts and facts about individual, social group, 
community, and urban environment, and hence is an important task in both 
artificial intelligence and mobile computing. In this dissertation, we study 
how to build effective analytical and predictive models of human behavior 
contexts and urban dynamics, which is crucial in building context-aware 
ubiquitous systems and enabling smart-city applications.

In the first part, we design effective learning methods to capture semantic 
behavior contexts from human mobile data, including personal mobility data and 
social interaction data. Our modeling methods contrast with existing approaches 
in that we address several typical challenges in real world mobile data 
jointly, including the presence of noise and data sparsity, and the absence of 
semantic labels, as well as enable exploratory, inference, and predictive 
purposes in a unified framework. For personal mobility data, we first design a 
Bayesian network to discover the routine behavior pattern from a single user's 
time-stamped mobility traces. Based on such, we then make non-trivial 
extensions
to address the difficult problem of representing mobility habits of multiple 
individuals in a unified way. To address the key challenge that multiple 
individuals rarely have spatial overlap or social connections in their 
mobility, we leverage the observation that temporal structures in habits can be 
highly shared across individuals. We design two methods based on novel 
extensions of matrix factorization and hierarchical Dirichlet processes to 
realize such population mobility models, and demonstrate how they help solve 
several challenging pervasive tasks, including routine behavior pattern 
discovery from sparse mobility data, individual mobility prediction under 
cold-start condition, and organizational rhythm discovery. We also apply 
similar methods to Bluetooth proximity data to discover social circles 
semantics for human social behavior characterization and social event 
prediction.

In the second part, we design effective methods to learn road latent cost from 
historical taxi trajectory data for urban sensing. Road latent cost quantifies 
how desirable each road is for traveling, and is a good representation of 
driving experience and urban dynamics. We first study how to robustly estimate 
the temporal dynamics of road travel time cost under temporal sparsity by 
exploiting temporal smoothness in a multi-task regression framework. This 
contrasts with existing work which either ignores such temporal dynamics or 
assumes it as a known function. In addition to travel time, a plenty of other 
hidden factors may influence the desirability of a road, but are impossible to 
obtain in practice. To address this, we propose to learn road latent cost from 
entire trajectories by modeling drivers' routing decisions based on inverse
reinforcement learning, while at the same time properly considering the 
heterogeneity of destinations so that trajectories with different goals can be 
learned jointly. In addition, observing that real trajectory data often 
contains a few anomalous trajectories, we design a sparse noise-oriented robust 
inverse reinforcement learning framework which can automatically identify and 
remove anomalies in cost learning. Real data experiments show that compared 
with past edge-centric approaches and noise-indifferent approaches, the road 
latent costs learned in our way are more useful and robust in facilitating 
typical smart-city applications, and require less data for learning.


Date:			Friday, 13 February 2015

Time:			3:00pm - 5:00pm

Venue:			Room 3494
 			Lifts 25/26

Chairman:		Prof. Howard Luong (ECE)

Committee Members:	Prof. Lionel Ni (Supervisor)
 			Prof. Qiong Luo
 			Prof. Raymond Wong
 			Prof. Jian-Ping Gan (MATH)
 			Prof. Jiannong Cao (Computing, PolyU)


**** ALL are Welcome ****