Output details

A principled approach to network-based classification and data representation

<24> Machine learning approaches. The originality of the paper is to propose the first rigorous methodology for mapping non-linear probabilistic classifiers into instance-based classifiers. Previous approaches used generative models with Fisher Information (FI) calculated in the space of model parameters, whereas we show a direct implementation with discriminative classifiers, by calculating the FI directly in data space. The theoretical significance of the paper is to explicitly define the local similarity measures implicit in the machine learning classifier. Its practical significance is to rigorously implement machine learning classifiers as methods for intelligent data access.