Published Paper Details:

The paper reports the results of the analysis of large-scale study data of diabetes prediction by employing an array of machine learning models. The Random Forest model is found to be highly discriminatory on the training set if its Area Under the Curve (AUC) value is close to one. Overfitting or erroneous classification threshold may be one of the probable issues as it has low accuracy in properly tagging data. The research also contrasted the Gaussian Naive Bayes, XgBoost, CatBoost, Gradient Boosting, and Logistic Regression models. Consistent performance of logistic regression on datasets with mid-level accuracy and AUC revealed equally well-balanced capability in classification and ranking. XgBoost and CatBoost did well in generalization and test data accuracy, and Gaussian Naive Bayes did well on the training data but suffered a dramatic decline in performance when executed on test and unseen data, possibly due to an overfitting problem. Gradient boosting also had a very close margin of accuracy when run on unseen data but had excellent discriminative capability on all the training, test, and unseen data and excellent generalization from the training data. The research ended with an agreement that which model would be used would be based on whether the specific requirement of the application was class discrimination, label prediction, or both being more salient.