Task 6 aims to learn comprehensive descriptive and predictive models of city traffic by combining multi-modal traffic views with situational context data given by external sources. Illustrative sources of context data to be considered include: 1) public events (such as festivals, concerts, conventions, sport events, etc.); 2) weather records and forecasts; 3) road incidents and other construction works; and, amongst others, 4) city ornament such as maps of schools, parks, commercial areas and other buildings of interest.

The integrative analysis of traffic data against their situational context offers unique opportunities to understand and anticipate traffic dynamics. In this context, Task 6 will discover statistically significant correlations between mobility dynamics and their situational context; and use them to enhance the targeted predictive models.

Two major activities will be accomplished:

A1) Comprehensive analysis of correlations against context data

In the context of Task 3, traffic records and context data were consistently integrated within a multi-dimensional database defined by a set of dimensions and (one or more) fact table. The entries in a given fact table (capturing traffic records) contain a key for each dimension (capturing context events linked to a specific date and location). Given such multi-dimensional database, this activity aims to find comprehensive correlations between dimensions and the entries of a given central fact.

In this context, this activity will explore two major tasks: 1) multi-dimensional subspace clustering, and 2) relational pattern mining. 

A2) Context-sensitive predictive models of mobility 

The established correlations in previous activity will be further considered to turn the predictive models of traffic sensitive to relevant context data.
Illustrating, scheduled events, construction works and weather predictions can be used to affect the anticipation of emergency occurrences in accordance with past correlation evidence. 

To this end, we will explore supervised learning methods well-prepared to learn from multi-dimensional data, with particular focus on the localization and classification of (motorized and non-motorized) traffic flow problems. In addition, this activity will establish a rigorous statistical ground to assess and improve the lower and upper bounds on predictions given the potentially high impact on mobility decisions.

This task builds upon public data sources of city situational context gathered in Task 2 and consolidated with urban mobility data in Task 3. The descriptive and predictive models of city traffic developed in Tasks 4 and 5 are enhanced here. The resulting context-sensitive statistical models will form an important basis for the contributions of Task 7. 

The INESC-ID, closely aided by LNEC and CML, will be responsible for the development and of the proposed algorithms given their extensive expertise on integrative learning from heterogeneous data (Henriques, 2015a). BPD-1 and BPD-2 grant holders will also work in this task.