Journal Article A Deep Neural Network for Pairwise Classification : Enabling Feature Conjunctions and Ensuring Symmetry

Atarashi, Kyohei  ,  Oyama, Satoshi  ,  Kurihara, Masahito  ,  Furudo, Kazune

10234pp.83 - 95 , 2017 , Springer
Pairwise classification is a computational problem to determine whether a given ordered pair of objects satisfies a binary relation R which is specified implicitly by a set of training data used for ‘learning’ R. It is an important component for entity resolution, network link prediction, protein-protein interaction prediction, and so on. Although deep neural networks (DNNs) outperform other methods in many tasks and have thus attracted the attention of machine learning researchers, there have been few studies of applying a DNN to pairwise classification. Important properties of pairwise classification include using feature conjunctions across examples. Also, it is known that making the classifier invariant to the data order is an important property in applications with a symmetric relation R, including those applications mentioned above. We first show that a simple DNN with fully connected layers cannot satisfy these properties and then present a pairwise DNN satisfying these properties. As an example of pairwise classification, we use the author matching problem, which is the problem of determining whether two author names in different bibliographic data sources refer to the same person. We show that the method using our model outperforms methods using a support vector machine and simple DNNs.
Advances in Knowledge Discovery and Data Mining : 21st Pacific-Asia Conference, PAKDD 2017, Jeju, South Korea, May 23-26, 2017, Proceedings, Part I, ISBN: 978-3319574530

Number of accesses :  

Other information