Semantic Signatures for Places of Interest

Search was always spatial. Searching and ranking text documents, for instance, is typically based on vector space models where similarity is calculated as the cosine of the angle between the term vectors. According to information foraging theory, such documents, say Web sites and their pages, contain information patches and their spatio-temporal properties are exploited by informavores to select a promising path based on the information scent [1]. In a broader sense, even typing in a search query, e.g., a sequence of keywords, is spatial. Terms closer to another are more likely to form meaningful n-grams. Of course, there is also spatial variation in what is being searched. A user living in a region with good public transit is more likely to search for bus routes and time tables [2]. Changing perspective, the dissemination and diffusion of information is also known to follow spatio-temporal patterns [3]. Finally, geo-fencing takes the act of querying out of the loop by pushing notifications to a device that enters or leaves a store, event, or another digitally bounded area. Examples range from sending coupons to users that walk by a store or alerting users about potential theft if their car leaves the parking area without them. In most cases, however, when we refer to spatial search we mean the fact that the user's geographic location provides important contextual cues to improve the relevance ranking between the query and the objects under consideration, e.g., Places Of Interest (POI). Spatial contextual awareness is just one of many contextual cues; other examples include the user's profile, navigation history, device type, and so forth. However, location is widely considered to be highly indicative of the user's intent [4]. Simplifying, a search engine will return a nearby coffee shop when queried for coffee instead of a more distant one or an Web page about the history or politics of coffee. What was true once only for the search on mobile devices is now also common practice for the desktop. Additionally, some systems support simple constraints on pre-defined attributes, e.g., place type, wheelchair access, wifi availability, or excluding localities that are currently closed based on the time the query is posed. A new category of applications centered around the idea of intelligent personal (digital) assistants relates events to locations, e.g., by showing traffic data while the user is on the way to work. Summing up, location matters for search, is typically …