Seeing a home away from the home: Distilling proto-neighborhoods from incidental data with Latent Topic Modeling

We report our initial experiments on a computational approach to neighborhood detection from community-authored location annotated content. Automated discerning of context from the physical environment is an important challenge in the age of mobile computing, where context awareness can often be a critical component in the design of pervasive applications. Furthermore, defining the context of a location, or a “neighborhood,” in an urban landscape is often a difficult task, since neighborhoods are often defined by intangible qualities such as shared function or community structure, rather than the prescribed geopolitical jurisdictions. In this study we seek to develop a data-driven definition of a neighborhood using topic modeling and socially annotated location data. We treat the question as a matter of latent topic discovery. Using incidental, geo-coded data gathered from locationsharing social networks, we show that we can discover canonical neighborhoods by analyzing co-occurence patterns in place categories. We report a qualitative evaluation of our approach, and interpret the validity of the discovered prototypical neighborhoods. We also discus future work and potential implications of our approach to the social sciences.