Authors:
Diogo Machado
1
;
2
;
Vítor Santos Costa
3
;
2
and
Pedro Brandão
1
;
2
Affiliations:
1
Instituto de Telecomunicações, Portugal
;
2
Faculty of Science of the University of Porto, Portugal
;
3
INESC-TEC, Portugal
Keyword(s):
Data Mining, Diabetes, Data-Balance, over-Sampling, Under-Sampling.
Abstract:
Imbalanced data sets pose a complex problem in data mining. Health related data sets, where the positive class is connected to the existence of an anomaly, are prone to be imbalanced. Data related to diabetes management follows this trend. In the case of diabetes, patients avoid situations of hypo/hyperglycaemia, which is the anomaly we want to detect. The use of balancing methods can provide more examples of the minority class, and assist the classifier by clearing the decision boundary. Nevertheless, each over-sampling and under-sampling method can affect the data set uniquely, which will influence the classifier’s performance. In this work, the authors studied the impact of the most known data-balancing methods applied to the Ohio and St. Louis diabetes related data sets. The best and most robust approach was the use of ENN with SMOTE. This hybrid method produced significant performance gains on all the performed tests. ENN in particular had a meaningful impact on all the tests. G
iven the limited volume of glycaemia-based data available for diabetes management, over-sampling methods would be expected to have a greater role in improving the classifier’s performance. In our experiments, the clearing of noise values by the under-sampling methods, produced better results.
(More)