Predictive analytics for human mobility and time-space behavior

This project focuses on the topic of human mobility and time-space behavior predictability based on Big Data. Human mobility research has evolved rapidly in the last years justified by a growing interest in predicting human time-space behavior for scientific, business and policy reasons. On the one hand, human predictability, even limited to the time-space domain, is a topic of great theoretical interest with profound ethical implications. Achieving some degree of collective predictability also enables multiple practical applications in business and policy making, for addressing problems in transportation, energy distribution or health care. One of the main factors underlying the growing interest in the subject is the increasing availability of proxy data of human mobility, such as telecom records, mobile application data, financial transactions, but also user-generated content that indirectly captures future intentions recorded through social media.
Most literature focuses on predictability based on long time-series of observations and on data directly linked to the phenomena at hand (e.g. GPS traces); it is developed around strong assumptions (e.g. stationary of mobility sequences, for instance generated by commuters) or uses domain knowledge to better model behavior (e.g. knowledge of travel mode options). Increasingly,  studies have started to focus on short, volatile sequences (e.g. those generated by tourists during short visits to a certain location) and attempt to exploit the massive amounts of data available to predict human behavior with very short or sparse data history. Additionally, a growing number of methods attempt to create predictors without ex-ante behavior assumptions, leading to agnostic forecasters based purely on data features.

This PhD will focus on telecom data and user generated data to produce forecasters of human mobility at various time and spatial scales. Seeking for maximum generality, it will focus on data-driven forecasts derived without previous behavior assumptions.

Research questions addressed in this PhD include:

•    Which classes of human mobility behavior can be predicted by which methods and under which assumptions?
•    What are the performances of methods such as Neural Networks, Expert Ensembles, or Self-Organizing Maps for short and long-term predictions? How well do they perform with different data?
•    Which type of policy questions do these methods answer to?
•    Which data characteristics, processing, integration or blending increase/decrease predictability of human behavior?