EGE12TH INTERNATIONAL CONFERENCE ON APPLIED SCIENCES
EFFECT OF BOOTSTRAP SAMPLING METHOD ON CLASSIFICATION PERFORMANCE IN DIABETES DATASET
Yayıncı:
Academy Global Publishing House
The success of machine learning models is directly related to the quality and quantity of the training dataset. However, data collection in the health field is usually a high-cost and challenging process. This necessitates working with limited data. Methods that provide data augmentation stand out as effective tools to overcome this limitation. The bootstrap sampling method aims to create larger datasets by performing random sampling with replacement from existing data and to increase model performance. In this study, the effect of applying the bootstrap sampling method at different scales on the diabetes dataset on the XGBoost classification model was investigated. The dataset used in the study includes various clinical characteristics of diabetic patients. The bootstrap sampling method was applied to create three different scenarios by resampling the dataset: (1) the original dataset, (2) the dataset doubled with bootstrap, and (3) the dataset tripled with bootstrap. The XGBoost algorithm was applied with the same hyperparameters in all scenarios. The datasets were divided into 70% training and 30% testing; performance evaluation was made on accuracy, sensitivity, selectivity, positive predictive value, negative predictive value and F1-score metrics. In addition, a complexity matrix was created for each scenario and the results were visualized in detail. In the original dataset, the accuracy rate was 73.59%, sensitivity 73.59%, selectivity 72.16%, F1-score 73.89%, positive predictive value (PPV) 71.18% and negative predictive value (NPV) 71.18%. In the case of doubling the dataset with the bootstrap method, accuracy was obtained as 95.88%, sensitivity 95.88%, selectivity 95.28%, F1-score 95.87%, PPV 95.67% and NPV 95.67%. It was observed that when the dataset was tripled with the bootstrap method, the accuracy was 96.39%, the sensitivity was 96.39%, the selectivity was 96.19%, the F1-score was 96.39%, the PPV was 95.95% and the NPV was 95.95%. The bootstrap method provided a consistent performance increase in all metrics. The bootstrap sampling method provided significant improvements in the performance of the classification model. Doubling the dataset resulted in a 19% increase in accuracy, while tripling it resulted in a 23% increase in accuracy. These findings emphasize that the bootstrap method is an effective strategy, especially in applications in the healthcare field where limited data is used.
