The nature of dataset set is the main factor that plays a role in that, but what are the other factors, Is the type of classifiers or the parameter related.
For the fixed classifier, try different classifier parameters. For example, if you use SVM as classification method, different parameter selection could result in huge difference in the ROC value.