A probabilistic model for entity disambiguation using relationships

TLDR: In this paper it is argued a better solution exist which analyzes not only features but also relationships, and standard feature-based data cleaning approaches can be employed.

Abstract: Graphs representing relationships among sets of entities are of increasing focus of interest in the context of data analysis applications. These graphs are typically constructed from existing datasets from which entities and relationships are extracted. For some of the entities, values in certain attributes would refer to other entities – such references determine relationships. Often, for certain datasets such references are given in the form of (string) descriptions. Each such description alone may not uniquely identify one entity as it is supposed to, but rather can match descriptions of multiple entities. Such cases are especially common if the datasets are collected not from one but multiple heterogeneous sources. Thus the correct linking of entities via relationships can be a nontrivial challenge which, if done incorrectly, can in turn impede further graph-based analyses. To overcome this problem, standard feature-based data cleaning approaches can be employed. In this paper we argue a better solution exist which analyzes not only features but also relationships.