多组学联合缺失数据填补方法的评价

发布时间：2018-05-14 11:53

本文选题：多组学数据 + 块缺失　；参考：《中国卫生统计》2017年04期

【摘要】：目的本研究旨在评价不同平台间"块缺失"数据的填补方法。如何在保证方差-协方差结构相对稳定的前提下提高多组学数据填补的精确度,对于后期数据挖掘有重要的意义。方法利用癌症基因组图谱(TCGA)数据库的肺癌数据(甲基化数据、基因表达数据),构建不同缺失比例的数据集(缺失比例分别为5%、20%、35%、50%和65%)。采用统计学填补方法均值法,马尔科夫蒙特卡洛法(MCMC)和机器学习填补法[邻近法(kNN),随机森林法(RF),多层感知机法(MLP)]对缺失数据进行填补,填补后数据集与原数据集进行比较。评价指标包括估计偏差和矩阵-2-范数。根据评价指标和填补时间,比较出填补效果最优、填补时间较短的方法。结果 MLP和kNN算法在各种缺失比例下均比其他填补方法有更优的效果,填补时间也相对较短。均值法的时间最短,在数据集缺失比例较小时(≤5%),填补效果与其他填补方法相当,但在高比例缺失情况下表现较差。在数据集高比例缺失情况下,RF和MCMC的填补效果优于均值法,但填补时间过长,不适用于实际工作。结论综合比较,机器学习填补方法中的MLP和kNN两法适合于甲基化数据和表达数据的填补。
[Abstract]:The purpose of this study is to evaluate the filling method of "block missing" data between different platforms. How to improve the accuracy of multi group data filling under the premise of ensuring the relative stability of variance covariance structure is of great significance for later data mining. Methods using the cancer data (TCGA) database of lung cancer data (methylation data, Gene expression data), set up data sets with different loss ratios (5%, 20%, 35%, 50% and 65%). Using statistical filling method mean method, Markoff Monte C Lofa (MCMC) and machine learning filling method (kNN), random forest method (RF), multi-layer perceptron (MLP)) fill the missing data and fill the post data. The set is compared with the original data set. The evaluation index includes the estimation deviation and the matrix -2- norm. According to the evaluation index and the filling time, it compares the method of filling the best effect and filling the short time. Results MLP and kNN algorithm have better effect than other filling methods under various missing proportions. The filling time is relatively short. Mean method is also relatively short. The time is the shortest, the proportion of missing data sets is small (less than 5%). The filling effect is equivalent to the other filling methods, but the performance is poor in the case of high proportion missing. In the case of high proportion missing data set, the filling effect of RF and MCMC is better than the mean method, but the filling time is too long and does not apply to the actual work. Conclusion comprehensive comparison, machine learning fill The two methods of MLP and kNN are suitable for the filling of methylation data and expression data.

【作者单位】：南京医科大学公共卫生学院生物统计学系;
【基金】：国家自然科学基金重点项目(81530088) 面上项目(81473070,81373102) 国家自然科学青年基金(81402764) 江苏省高校优势学科资助
【分类号】：O212.1;R73-3

【相似文献】