ISSN:
1573-7497
Keywords:
neural networks
;
structured objects
;
machine learning
;
classification
;
similarity
;
nearest neighbor
Source:
Springer Online Journal Archives 1860-2000
Topics:
Computer Science
Notes:
Abstract Labeled graphs are an appropriate and popular representation of structured objects in many domains. If the labels describe the properties of real world objects and their relations, finding the best match between two graphs turns out to be the weakly defined, NP-complete task of establishing a mapping between them that maps similar parts onto each other preserving as much as possible of their overall structural correspondence. In this paper, former approaches of structural matching and constraint relaxation by spreading activation in neural networks and the method of solving optimization tasks using Hopfield-style nets are combined. The approximate matching task is reformulated as the minimization of a quadratic energy function. The design of the approach enables the user to change the parameters and the dynamics of the net so that knowledge about matching preferences is included easily and transparently. In the last section, some examples demonstrate the successful application of this approach in classification and learning in the domain of organic chemistry.
Type of Medium:
Electronic Resource
URL:
http://dx.doi.org/10.1023/A:1008320413168
Permalink