肿瘤参数属性偏序结构可视化实现乳腺癌诊断

为实现乳腺癌数据规则可视化,提出一种基于Lasso和增量学习结合的、以改进的属性偏序结构图为可视化工具的乳腺癌诊断规则提取方法。采用乳腺癌数据为数据源基础上算法分为4步:首先使用Lasso方法进行特征选择实现降维,在9个特征中选出前4个关联度最大的特征;其次进行基于Gini指数的连续数据粒化,通过增量学习方式动态生成形式背景;再次融合二次Lasso筛选,将维数由17降为3;最后使用新的基于基尼指数和覆盖对象的行列优化方法生成属性偏序结构图可视化规则,提取出规则7条。将数据处理结果与主流分类器对比,结果表明,基于该算法的规则提取实现96.52%的诊断准确率,均高于随机森林(94.25%)、Adaboost(90.00%)、1NN(91.33%)、3NN(90.67%)、支持向量机算法(95.00%)。最后采用不同增量比例(10%~90%)数据验证增量学习算法效果,表明顺序学习数据量达到30%时模式已经完备,数据量在20%时准确率已经接近支持向量机算法,证明该方法是一种用于诊断可视化的规则发现的有效手段。

Diagnosis of Breast Cancer Based on Tumor Parameters and Visualization of the Attribute Partial Order Structure Diagram

In order to realize the visualization of the rules of breast cancer data, a method based on the combination of Lasso and incremental learning, was proposed, using the optimized attribute partial order structure diagram as a tool. Firstly, having the dimensions reduced by using Lasso to select the features of the breast cancer data, and four attributes that gained the largest correlation were selected from nine features. Granulation process was completed under the Gini index, generating the formal context by means of the incremental learning algorithm. Next, the second Lasso process was completed, which made the dimensions reduced from 17 to 3. Meanwhile, a new method processing the rows and columns of the formal context based on the Gini index and the covering theory was proposed to generate the attribute partial order structure diagram to visualize the rules concerned. As there have been seven rules extracted by analyzing the diagram reported in literatures,we compared the proposed classification accuracy of the method with those classical mainstream classifiers. Results showed that the classification precision of our method reached 96.52%, higher than the other five classifiers including Random Forest (94.25%), Adaboost (90.00%), 1NN (91.33%), 3NN (90.67%), and SVM (95.00%). At last, different incremental proportional (10%-90%) data were used to verify the effect of incremental learning algorithm, results showed that the model had been completed when the amount of data reached 30%, and the precision was almost approaching to that of support vector machine, which proved that the proposed method represented an effective means of visualizing the diagnosis rules of breast cancer.