A multiobjective optimization based entity matching technique for bibliographic databases. (15th December 2016)

• Record Type: Journal Article
• Authors: Mishra, Sumit; Saha, Sriparna; Mondal, Samrat
• Is Part Of: Expert systems with applications. Volume 65(2016)
• Page Start: 100
• Page End: 115
• Publication Date: 2016-12-15
• Subjects: Bibliographic database -- Entity matching -- Multiobjective optimization -- Genetic algorithm -- Elitism -- Pareto-optimal solutions; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2016.07.043
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Multiobjective based solution framework for entity matching. Bibliographic database is used. Automatic determination of number of clusters and partitions. New encoding strategy, new mutation operators are explored. Several distance measures and cluster validity indices are used. Abstract: With the increasing use of on-line resources, the size of the bibliographic database is growing day by day. The available huge amount of data belong to various entities. It is difficult to automatically identify the records which belong to a particular entity. Mapping the records to the corresponding entity is termed as the entity matching problem. In bibliographic database many attributes change over time. For example - affiliation of an author changes frequently. Many authors generally use different email-ids. The names of co-authors also change with time. All these aspects have made the entity matching problem challenging. Generally an entity matching task is carried out by constructing a feature vector to represent a record, then a classifier is trained to classify each feature vector. But for bibliographic database it is very difficult and time consuming to generate some manually annotated labeled data to train a classifier. Inspired by this observation, we have proposed an unsupervised approach for entity matching problem using non-dominated sorting genetic algorithm-II (NSGA-II). A new encoding strategy is used to encode the clusters in the form of a chromosome. New mutation and crossover operators are proposed which are suitable for bibliographic data clustering. Different distance measures are used to measure the dissimilarities between records. Finally, solutions are evolved using the search capability of NSGA-II. Experimental evaluations are carried out with 247 different combinations of eight objective functions for eight different bibliographic datasets. A comparative analysis with two existing systems - DBLP and ArnetMiner, shows that the proposed technique can produce better results in many cases.