On the one hand, classification applications modelled by structural pattern recognition, in which elements are represented as strings, trees or graphs, have been used for the last thirty years. In these models, structural distances are modelled as the correspondence (also called matching or labelling) between all the local elements (for instance nodes or edges) that generates the minimum sum of local distances. On the other hand, the generalised median is a well-known concept used to obtain a reliable prototype of data such as strings, graphs and data clusters. Recently, the structural distance and the generalised median has been put together to define a generalise median of matchings to solve some classification and learning applications. In this paper, we present an improvement in which the Correspondence edit distance is used instead of the classical Hamming distance. Experimental validation shows that the new approach obtains better results in reasonable runtime compared to other median calculation strategies.
MORENO-GARCÍA, C.F. and SERRATOSA, F. 2018. Modelling the generalised median correspondence through an edit distance. In Bai, X., Hancock, E.R., Ho, T.K., Wilson, R.C., Biggio, B. and Robles-Kelly, A. (eds.) Structural, syntactic, and statistical pattern recognition: proceedings of the 2018 Joint International Association for Pattern Recognition (IAPR) international workshops on structural and syntactic pattern recognition (SSPR 2018), and statistical techniques in pattern recognition (SPR 2018) (S+SSPR 2018), 17-19 August 2018, Beijing, China. Lecture notes in computer science, 11004. Cham: Springer [online], pages 271-281. Available from: https://doi.org/10.1007/978-3-319-97785-0_26