Efficiently Finding Similar Objects on Ontologies Using Earth Mover’s Distance

Abstract

Ontologies are being progressively used to capture the semantics of information from various sources. They have wide area of usage ranging from artificial intelligence, natural language processing to web content and biology. This paper proposes the problem of finding similar objects that have been defined as a set of terms from an ontology. We consider tree-based ontologies where a node represents a term and an edge weight defines the distance or dissimilarity between corresponding terms. For object distance, Earth Mover’s Distance (EMD) is used as it outperforms other distance measures like average and minimum pairwise distance. EMD, however is highly computationally intensive as it involves solution to linear programming (LP) problem. We propose an efficient lower bound on computing EMD by aggregating the terms in the ontology at the first level of the tree. This reduces the number of terms, thereby decreasing the number of flow variables and making it computationally faster. Range queries that use the lower bound runs faster by up to a factor of 20, as approximately 97% percentage of database objects are pruned, thereby saving expensive EMD calculations.