Mining Heterogeneous Urban Data at Multiple Granularity Layers

The recent development of urban areas and of the new advanced services supported by digital technologies has generated big challenges for people and city administrators, like air pollution, high energy consumption, traffic congestion, management of public events. Moreover, understanding the perception of citizens about the provided services and other relevant topics can help devising targeted actions in the management. With the large diffusion of sensing technologies and user devices, the capability to generate data of public interest within the urban area has rapidly grown. For instance, different sensors networks deployed in the urban area allow collecting a variety of data useful to characterize several aspects of the urban environment. The huge amount of data produced by different types of devices and applications brings a rich knowledge about the urban context. Mining big urban data can provide decision makers with knowledge useful to tackle the aforementioned challenges for a smart and sustainable administration of urban spaces. However, the high volume and heterogeneity of data increase the complexity of the analysis. Moreover, different sources provide data with different spatial and temporal references. The extraction of significant information from such diverse kinds of data depends also on how they are integrated, hence alternative data representations and efficient processing technologies are required. The PhD research activity presented in this thesis was aimed at tackling these issues. Indeed, the thesis deals with the analysis of big heterogeneous data in smart city scenarios, by means of new data mining techniques and algorithms, to study the nature of urban related processes. The problem is addressed focusing on both infrastructural and algorithmic layers. In the first layer, the thesis proposes the enhancement of the current leading techniques for the storage and elaboration of Big Data. The integration with novel computing platforms is also considered to support parallelization of tasks, tackling the issue of automatic scaling of resources. At algorithmic layer, the research activity aimed at innovating current data mining algorithms, by adapting them to novel Big Data architectures and to Cloud computing environments. Such algorithms have been applied to various classes of urban data, in order to discover hidden but important information to support the optimization of the related processes. This research activity focused on the development of a distributed framework to automatically aggregate heterogeneous data at multiple temporal and spatial granularities and to apply different data mining techniques. Parallel computations are performed according to the MapReduce paradigm and exploiting in-memory computing to reach near-linear computational scalability. By exploring manifold data resolutions in a relatively short time, several additional patterns of data can be discovered, allowing to further enrich the description of urban processes. Such framework is suitably applied to different use cases, where many types of data are used to provide insightful descriptive and predictive analyses. In particular, the PhD activity addressed two main issues in the context of urban data mining: the evaluation of buildings energy efficiency from different energy-related data and the characterization of people's perception and interest about different topics from user-generated content on social networks. For each use case within the considered applications, a specific architectural solution was designed to obtain meaningful and actionable results and to optimize the computational performance and scalability of algorithms, which were extensively validated through experimental tests.