| Summary: | 碩士 === 逢甲大學 === 資訊工程所 === 92 === The k-nearest neighbor rule (k-NNR) is commonly used in applications of classifiers and data mining and the related area due to its simplicity and effectiveness. Theoretically, the goal of designing an optimal k-NNR classifier is to maximize the classification accuracy while minimizing the sizes of both the reference and feature sets. Recently, some studies tackled the multi-objective function by using the weighted-sum approach. However, such approaches are often criticized on its robustness, because they are sensitive to the weight values, and the weight values are usually subject to the practitioner’s decision.
In order to avoid those drawbacks, an approach using intelligent multi-objective evolutionary algorithm (IMOEA) is applied to design optimal multi-objective k-NNR classifiers in this paper. The advantages of IMOEA are as follows: (1) The generalized Pareto-based scale-independent fitness function can assign discriminative fitness value to individuals based on the Pareto concept. (2) IMOEA can converge to high quality solutions by making use of the system reasoning ability of intelligent gene colloector. (3) IMOEA incorporates the elitism strategy to speed-up the convergence.
In this paper, the proposed apprach is compared with other approaches by using eleven data sets obtained of UCI machine learning database. The experimental results show that, IMOEA-designed 1-nn classifier outperforms SPEA-designed 1-nn classifier in terms of classification accuracy, the number of reference set and the number of feature set. The non-dominated solutions obtained by IMOEA dominate 88.98% of non-dominated solutions obtained by SPEA, averaged from 11 data sets. Considering only classification accuracy, IMOEA-designed classifier outperforms C4.5 at a ratio of 5.95%, averaged from 11 data sets.
|
|---|
