In this talk we present two historical models of human migration from the 19th and 20th centuries, and discuss how they apply to location data on the Web, in the 21st Century.
Record linkage is the task of identifying which records in one or more data collections refer to the same entity, and address is one of the most commonly used fields in databases. Hence, segmentation of the raw addresses into a set of semantic fields is the primary step in this task. In this paper, we present a probabilistic address parsing system based on the Hidden Markov Model. We also introduce several novel approaches of synthetic training data generation to build robust models for noisy real-world addresses, obtaining 95.6% F-measure. Furthermore, we demonstrate the viability and efficiency of this system for large-scale data by scaling it up to parse billions of addresses.
This paper is concerned with the automatic prediction of the zoom level at which to present a map with results for an informal location description. We propose the use of identifiability (relative uniqueness) and zoom level of each component geospatial expression (GE) in the location description, as a means of predicting the appropriate zoom level for the overall description. We apply a simple classification approach to zoom level prediction, and compare results using gold-standard and automatically-inferred GE information. We find the approach to have strong promise, including relative to the zoom level used in results from Google Maps for our location descriptions dataset.
Location nowadays is an important aspect of the Web. One scenario in this respect are archives or collections of geo-tagged media items. More concretely, we can think of collections in the arts and humanities available via OAI-PMH (a protocol for metadata harvesting) on the web or web accessible personal media archives maintained in a peer-to-peer manner. In such scenarios for search problems, source selection becomes an important aspect. For example, we would like to access only those collections containing media items in a certain geospatial region (maybe we are interested in images from Shanghai only). Here, the geospatial search criterion allows for a high selectivity. What is needed in such a scenario are expressive and nevertheless compact representations or descriptions of the ``geospatial footprint'' of each collection. A minimum bounding rectangle would be a trivial but not very accurate option. Generally, summarization techniques for this purpose can be distinguished into three categories, geometric approaches, space partitioning approaches and hybrid approaches. In this work, we present novel hybrid techniques, which mostly apply a set of approximating minimum area rectangles for subspace description together with quantization techniques in order to increase the selectivity of the summaries and, at the same time, keep the storage requirements small.
Indoor positioning system (IPS) identifies positions of various indoor objects, and is a key technology to achieve sophisticated Indoor Location-Aware Services (InLAS). In most conventional systems, InLAS and IPS are tightly coupled. That is, one system does not supposed to reuse indoor location data and program of another system. This makes individual systems complex and difficult to manage. To cope with the problem, we propose Data Model for Indoor Location (DM4InL), which prescribes a common data schema, independent of implementation of IPS or the usage of InLAS. The proposed DM4InL represents the location of every indoor object in a standard way, by using three kinds of models: location, building and object models. We also design the fundamental API, which implements typical queries to the indoor location data from external applications. The proposed method achieves loose-coupling of InLAS and IPS, which significantly improves the efficiency and reusability in the InLAS development.
