The 12th International Workshop on Urban Computing

The goals and framework of urban computing result in four folds of challenges in the context of data mining:

• Adapt machine learning algorithms to spatial and spatio-temporal data: Spatio-temporal data has unique properties, consisting of spatial distance, spatial hierarchy, temporal smoothness, period and trend, as compared to image and text data. How to adapt existing machine learning algorithms to deal with spatio-temporal properties remains a challenge.

• Combine machine learning algorithms with database techniques: Machine learning and databases are two distinct fields in computing science, having their own communities and conferences. While people from these two communities barely talk to each other, we do need the knowledge from both sides when designing data analytic methods for urban computing. The combination is also imperative for other big data projects. It is a challenging task for people from both communities to design effective and efficient data analytics methods that seamlessly and organically integrate the knowledge of databases and machine learning.

• Cross-domain knowledge fusion methods: While fusing knowledge from multiple disparate datasets is imperative in a big data project, cross-domain data fusion is a non-trivial task given the following reasons. First, simply concatenating features extracted from different datasets into a single feature vector may compromise the performance of a task, as different data sources may have very different feature spaces, distributions and levels of significance. Second, the more types of data involved in a task, the more likely we could encounter a data scarce problem. For example, five data sources, consisting of traffic, meteorology, POIs, road networks, and air quality readings, are used to predict the fine-grained air quality throughout a city. When trying to apply this method to other cities, however, we would find that many cities cannot find enough data in each domain (e.g. do not have enough monitoring stations to generate air quality data), or may even not have the data of a domain (like traffic data) at all.

• Interactive visual data analytics: Data visualization is not solely about displaying raw data and presenting results, though the two are general motivation of using visualization. Interactive visual data analytics becomes even more important in urban computing, seamlessly combining visualization methods with data mining algorithms as well as a deployment of the integration on a cloud computing platform. It is also an approach to the combination of human intelligence with machine intelligence. The interactive visual data analytics also empower people to integrate domain knowledge (such as urban planning) with data science, enabling domain experts to work with data scientist on solving a real problem in cities.