7. Acknowledgment 8. References Cervical Cell Data 6. Concluding Remarks

Non-parametric decision rules, such as the nearest neighbor (NN) rule, are attractive because no a priori knowledge is required concerning the underlying distributions of the data. Two traditional criticisms directed at the NN-rule concern the large amounts of storage and computation involved due to the apparent necessity to store all the sample (training) data. Thus there has been considerable interest in “editing” or “thinning” the training data in an attempt to store only a fraction of it. Previous editing algorithms suffered from the drawback that they delivered edited sets that were not decision-boundary consistent, i.e., the decision boundary determined by the edited set differed from that specified by the entire original training data. In this paper several geometric methods based on proximity graphs are proposed for editing the training data for use in the NN-rule. Most notably, one of the methods yields a decision-boundary consistent edited set and therefore a decision rule that preserves all the desirable convergence properties of the NN-rule that is based on the original entire training data. The methods are all derived from the Voronoi diagram of the sample data and make use of subgraphs of the Delaunay triangulation. The methods are compared empirically through experiments on synthetic data as well as real world data in the automatic detection of cervical cancer. Finally, algorithms for the efficient implementation of these techniques are discussed.