Relational Classification Using Automatically Extracted Relations by Record Linkage

Relational classifiers often outperform traditional classifiers which assume that objects are independent. For applying relational classifiers relations are required. Since data often is noisy and unstructured, these relations are not given explicitly, but need to be extracted. In this paper we show a framework for relational classification that first automatically extracts relations from such a noisy database and then applies relational classifiers. For extracting relations we use techniques from the field of record linkage that learn the characteristics for a relation from pairwise features. With the proposed framework relational classifiers can be applied without requiring manual annotation.