Recommending spatial classes for entity interlinking in the web of data

Abstract

Recent advances in the web informatics domain bring closer the realization of Web of Data, a global interconnected data space where richer entity descriptions are easily retrievable and reusable. A key Web of Data component is the establishment of links between related entities. Link Discovery tools can be utilized for the (semi) automatic identification and linkage of related entities between a pair of entity sets. However, they require the manual examination and selection of Web of Data datasets (or sub parts of them) that will be used for link establishment. This research focuses on proposing automated methods, which search in Web of Data datasets and recommend pairs of classes that may contain related entities and thus can be used as input in Link Discovery tools. We approach the problem from a geographical perspective by exploiting the spatial information of classes i.e. the location of their instances. We intuitively believe that classes that present similar spatial distribution is likely to contain related entities. To achieve scalability at web scale, we study and implement spatial summarization methods that capture the spatial distribution of each class. To identify relevant classes, we investigate and propose techniques that act on the summaries to compute their similarity. We (a) evaluate two aspects of our methodology, namely the ability of identifying relevant classes effectively and performing at web scale efficiently and (b) compare our approach with other state of the art dataset recommendation for interlinking approaches.