When Similar Problems Don’t Have Similar Solutions

摘要: The performance of a Case-Based Reasoning system relies on the integrity its case base but in real life applications available data used to construct invariably contains erroneous, noisy cases. Automated removal these cases can improve accuracy. In addition, error rates for nearest neighbour classifiers often be reduced by removing give smoother decision boundaries between classes. this paper we argue that optimallevel boundary smoothing is domain dependent and, therefore, our approach reduction reacts characteristics set an appropriate level smoothing. We present novel, yet transparent algorithm, Threshold Error Reduction, which identifies and removes with aid local complexity measure. Evaluation results confirm it superior benchmark algorithms.