A Weighted Distance Learning Method for Graph Classification

A weighted distance-learning method is used, in this paper, to optimize the accuracy of Nearest-Neighbour classifier for a graph-based data set. This algorithm tunes the weights of the training instances to minimize an analytical approximation of the classification-error on the training data. With this method, not only the distance function, which has an effective role in the performance of classifier, is tuned, but also a prototype reduction is achieved as a side effect to reduce both of time and memory complexity of the classifier. This efficiency is more important in the case of having graph instances, because computing the distance of a query graph with each prototype graph is a time-consuming process due to isomorphism. By applying the proposed method on IAM, a benchmark graph dataset, the uniformly good behaviour of the proposed method is presented having results competitive with the related methods.