基于伪度量的案例推理改进算法

Improved Case Reasoning Algorithm Based on Pseudo-Metric

案例推理算法中案例之间的相似度非常关键,它影响着案例推理中最重要的两个部分:案例检索和案例重用。不断有研究尝试设计更合理的度量空间,结合神经网络等技术以改进算法关键阶段。该文采用伪度量作为案例的相似度度量,用神经网络拟合案例之间的度量函数及预测案例相似度,在案例重用阶段用新的公式取代以往的聚类方法,减少了构建匹配池阶段案例的匹配数量,解决以往算法重用阶段聚类方法带来信息过早定值化的问题。新设计由算法直接输出预测结果并判断目标案例的分类。实验验证了该算法在实验数据集上对比案例推理、SVM等准确率提升了2%,对比文献[9]的基于伪度量的案例推理算法平均运行时间减少到2.4%,且在正负样本不平衡数据上表现更优,优化了案例推理的过程。

The similarity between cases in case-based reasoning algorithm is critical,which affects the two most important parts of case-based reasoning:case retrieval and case reuse.There are constant research attempts to design a more reasonable metric space and improve the key stage of algorithm by combining neural network and other techniques.In this paper,pseudo-metrics are used as the similarity measures of the cases,and the neural network is used to fit the metric function between the cases and predict the similarity of the cases.In the stage of case reuse,the new formulas are used to replace the previous clustering methods,which reduces the number of matching cases in the stage of constructing matching pool,and solves the problem that the clustering method in the stage of case reuse brings premature information quantization.In the new design,the algorithm directly outputs the prediction results and judges the classification of the target case.Experiments verify that the accuracy of the proposed algorithm on the experimental data set is improved by 2%compared with case inference and SVM.Compared with the case-based reasoning algorithm based on pseudo-metrics in literature[9],the average running time is reduced to 2.4%.It performs better in the positive and negative sample unbalanced data,thus optimizing the process of case reasoning.