The study of under- and over-sampling methods’ utility in analysis of highly imbalanced data on osteoporosis

详细信息    查看全文

• 作者：M. Bach^a ; ^{malgorzata.bach@polsl.pl} ; ; A. Werner ; ^a ; ^{aleksandra.werner@polsl.pl} ; ; J. Żywiec^b ; ^{jzywiec@sum.edu.pl} ; ; W. Pluskiewicz^c ; ^{osteolesna@poczta.onet.pl} ;
• 关键词：Classification ; Imbalanced data ; Osteoporosis ; Performance measures ; Sampling methods
• 刊名：Information Sciences
• 出版年：2017
• 出版时间：April 2017
• 年：2017
• 卷：384
• 期：Complete
• 页码：174-190
• 全文大小：1300 K
• 卷排序：384

文摘

Osteoporosis is a frequent bone disease without typical early symptoms but with serious complications e.g. low-energy bone fractures. Patients with risk factors should be screened for proper diagnosis as early as possible. Unfortunately, the registered medical data are often highly imbalanced. That is why the machine-based data processing is difficult or even impossible. Considering this, our goal was to search for the best method of coping with the problem of imbalancing in relation to the analysed data regarding the osteoporotic patients. Therefore, we checked several paradigms of classifiers in synergy with preprocessing techniques to address the inner skewed class distribution of the data.In the source dataset 92.6% of instances related to patients without any fractures (negative cases) and only 7.41% to patients (positive cases) who reported at least one fracture. To alleviate class imbalance there were examined not only data-level methods which in fact modify the input dataset, but also ensemble ones that strengthen the results of the base algorithms. In the first group the under- and over-sampling methods were used, such as random undersampling, edited nearest neighbours and synthetic minority over-sampling techniques, while in the second one – a range of methods based on various subsets of training data were analysed. Also various combinations of the above mentioned were investigated. Additionally, we propose the way how to find the balancing level which, without excessive distortion of the input, raw data, will give the appropriate classification efficiency.The aim of our experiment was to identify which of an undersampling or an oversampling approach with reference to the simple and the ensemble-based classifiers allows to achieve the best results.The outcomes of the comparative studies concerning imbalancing problem with regard to our dataset showed that the highest efficiency was achieved while using the synthetic minority over-sampling technique and RandomForest classifier. As far as the optimal balancing level is concerned, we empirically determined that 300% oversampling with the synthetic minority over-sampling method combined with edited nearest neighbours undersampling allowed to gain the required precision of classification.