多分类器集成系统在卷烟感官评估中的应用

Application of multiple classifier systems in cigarette sensory evaluation

烟草化学成分与感官质量存在着某种对应关系,使用传统的单一分类器方法进行卷烟感官评估指标预测对数据样本依赖性强,且无法克服噪声影响。采用了k-NN(k=3)、k-NN(k=5)、C4.5、BPNN、SVM五种不同的分类方法进行卷烟感官质量的评估预测,并比较了它们的预测正确率,结果表明SVM和k-NN效果较好;同时用不同的采样方法和投票方法搭建了6种多分类器集成系统进行感官评估,实验可得多分类器集成可以起到比单一分类器更好的效果,其中基于训练表现的加权投票方法(不抽样)具有较高的正确率,对于指标光泽、香气、谐调、杂气、刺激和余味的预测准确率分别为90.63%、62.18%、97.20%、86.74%、74.17%和72.16%,比单一分类器中最好的结果分别改进0.69%、0.32%、0.28%、0.59%、0.52%和1.78%。

There exist some kinds of correlation between chemical components in tobacco and sensory quality of cigarette. The existing methods using single classifier to evaluate cigarette sensory quality depend largely on the property of data sample and cannot overcome the effect of data noise. Five different classifying methods, namely k-NN（k=3）, k-NN（k=5）, C4.5, BPNN, and SVM, were used to predict and evaluate sensory quality, and their prediction accuracy were compared. Results showed that SVM and k-NN had better effect. Moreover, six Multiple Classifier Systems（MCS） were built to make sensory evaluation using various sampling and voting methods. The experiments showed that integration of various classifiers could get better results than single classifier and the weighting-sum method based on training performance（no sampling） had better accuracy. The prediction accuracy of luster, aroma, harmony, offensive odor, irritation and aftertaste are 90.63%, 62.18%, 97.20%, 86.74%, 74.17% and 72.16%, 0.69%, 0.32%, 0.28%, 0.59%, 0.52% and 1.78% higher than that of single classifier respectively.