Supervised data mining in networks: Link prediction and applications

• Autores: Víctor Martínez Gómez
• Directores de la Tesis: Fernando Berzal (dir. tes.) , Juan Carlos Cubero Talavera (codir. tes.)
• Lectura: En la Universidad de Granada ( España ) en 2018
• Idioma: inglés
• ISBN: 9788491639824
• Número de páginas: 223
• Tribunal Calificador de la Tesis: Juan Miguel Medina Rodríguez (presid.) , Daniel Sánchez Fernández (secret.) , Ernesto Estrada (voc.) , Ernestina Menasalvas (voc.) , Santiago Escobar Román (voc.)
• Enlaces
  • Tesis en acceso abierto en: DIGIBUG
• Resumen

  • Link prediction is the problem of predicting the existence of currently-unobserved links or links that will appear in the future between pairs of nodes in complex networks. This problem has attracted a great deal of attention from researchers in diverse disciplines due to its applicability in a wide range of tasks, such as the identification of disease-associated candidate genes or the improvement of recommendations suggested by recommender systems. This PhD dissertation comprises different lines of work, all of them closely related to the link prediction problem.

    On the one hand, after an exhaustive study of the state of the art in link prediction, the main limitations of currently proposed approaches were identified. These limitations were related to the difficulties associated to the trade-off between scalability and performance in link prediction techniques. Two scalable link prediction techniques were proposed that follow different approaches to exploit local network features.

    On the other hand, different applications of link prediction techniques were addressed. We proposed a novel algorithm for generic prioritization, such as disease-gene prioritization, which achieved better results than other state-of-the-art techniques due to its capacity for integrating heterogeneous data sources. We also developed a novel algorithm for word sense disambiguation of semantic relations between concepts, based on link prediction and without the requirement of annotated data. We showed how our algorithm achieved better accuracy than other state-of-the-art techniques in different evaluation tasks and how relations extracted using our approach could improve the performance of state-of-the-art general-purpose word sense disambiguation techniques. In addition, since node role influences how links are formed in complex networks, we developed a novel distance metric based on the concept of automorphic equivalence with application to node role discovery.

    Finally, we developed a software framework for network data mining. This framework, called NOESIS, contains efficient implementations of an extensive list of network-related algorithms, including a complete library of link prediction techniques.