An authority-flow based ranking approach to discover potential novel associations between Linked Data

• Published in: Semantic Web Vol. 5; no. 1; pp. 23 - 46
• Main Authors: Vidal, María-Esther, Rivera, Jean-Carlo, Ibáñez, Luis-Daniel, Raschid, Louiqa, Palma, Guillermo, Rodriguez, Héctor, Ruckhaus, Edna
• Format: Journal Article
• Language: English
• Published: 2014
• Subjects: Bayesian analysis; Clouds; Drugs; Medical; Networks; Ranking; Sampling; Semantics
• ISSN: 1570-0844
• DOI: 10.3233/SW-130101

Under the umbrella of the Semantic Web, Linking Open Data projects have made available a large number of semantically intra- and inter-connected links. As an example, in the biomedical domain, data about disorders, disease related genes and proteins, clinical trials, and drugs or interventions are accessible on the Linked Open Data cloud. In addition, domain ontologies have been used to annotate scientific data. For instance, publications in PubMed have been annotated using controlled vocabulary (CV) terms from ontologies such as the Medical Subject Header (MeSH) or the Unified Medical Language System (UMLS). These annotations have been successfully mined to discover associations between drugs and diseases using techniques that have been labeled as Literature-Based Discovery (LBD). Given the large scale of the linked datasets in the Linked Open Data cloud, there is a need to develop scalable techniques that can provide answers in close to real time, to explain a phenomena, to identify anomalies, or to explore a discovery. This paper describes an authority flow based ranking technique that is inspired by LBD methods. The ranking is tailored to a layered graph. The input terms are in the first layer and the ranking will efficiently identify and assign high scores to terms in a third (or subsequent) layer, corresponding to potential novel discoveries. The terms, links and scores are modeled as a Bayesian network. Two sampling techniques are proposed to only traverse the terms that may have high scores. The first technique implements a Direct Sampling reasoning algorithm to approximate the ranking scores of nodes in the Bayesian network; it visits only the nodes with the highest probability. The second technique samples paths in the Bayesian network with the highest conditional probability. An experimental study reveals that the proposed ranking techniques are able to reproduce state-of-the-art discoveries. In addition, the sampling-based approaches are able to reduce execution times and reach high levels of accuracy.