基于多粒度一致覆盖约简的混合数据规则学习

Rule Learning for Hybrid Data Based on Multi-granulation Consensus Covering Reduction

许多已提出的规则提取算法不适用于混合数据,而混合数据广泛存在于实际应用中.针对上述问题,提出了一种基于多粒度一致覆盖约简的规则学习方法.首先,在不同粒度下导出每个样本的多粒度表示,从而获得所有粒度下每个样本的最大一致规则;其次,基于一致性的原则计算每个最大一致规则的权重;最后,提出了一种基于最大一致规则权重的覆盖约简算法以获得最小规则集.实验结果表明:所提出的算法能够有效地选择一组最小规则集,相比一些常见的规则提取技术,所提出的算法具有更好的分类性能.另外,所提算法在噪声扰动下具有更好的鲁棒性.

Most existing rule extraction algorithms are not applicable to hybrid data, which widely exist in real-world applications. To address this problem, we present a rule extraction technique based on multi-granulation consensus covering reduction. First, derive multi-granulation representation of each sample under different granularities, and obtain the maximal consistent region of each sample on all granularities. Then measure the weight of each maximal consistent region based on consensus prineiple. Finally, a covering reduction algorithm based on weight for all maximal consistent regions is presented to derive a minimal set of decision rules. Experimental results show that our method can effectively select a set of decision rules, and performs well in comparison with other popular rule extraction techniques. In addition, the proposed method is more robustness against noise for noisy data.