DEA (Data Envelopment Analysis) is a management science method that has been widely applied for performance analysis in various sectors. The application in this paper for treating congestion in DEA are extended by according them chance constrained programming formulations. A shortcoming of previous DEA applications is that it has been used to mainly evaluate ex ante performance.Congestion indicates an 1 Corresponding author:Farhad Hosseinzadeh Lotfi, E-mail:hosseinzadeh lotfi@yahoo.com 3170 F. Hosseinzadeh Lotfi et aleconomic state where inputs are overly invested. Evidence of congestion occurs whenever reducing some inputs can increase outputs. A congestion in stochastic DEA model is presented and then it is reformulated in the manner that the congestion in stochastic model incorporate future information. This research applies the approach to plan the restructure strategy of Iranian commercial banks.

The importance of predictor variables in every approach is obtained by sensitivity analysis. Accordingly, fasting blood sugar and 2-hour blood sugar (postprandial glucose) have been identified as two important variables in women. Glomerular filtration rate and mean arterial blood pressure in women are other variables having been identified in this research as predictors of the development of diabetes in women. In general, the results indicate that the variables of waist circumference and height are equally important in the incidence of diabetes, even ore than fasting glucose. In order to build predictive models, 6 main and conventional methods in data mining (decision trees with C5.0 algorithms, CART, QUEST, MLP algorithm neural network, support vector machine (SVM), Bayesian simple model (Naïve Bayes)) and a conventional model in epidemiological studies (logistic regression) have been used. The results showed that Naïve Bayes model is not sensitive to unbalanced data in both men and women so that its sensitivity in women with unbalanced data to total variables is 78%. Although other methods represented high characteristics of unbalanced data, they have very low sensitivity. Their average sensitivity to unbalanced data with total variables is 34.8% in women. The performance of all 6 methods of classification is comparable to balanced data; they are even better than logistic regression. In this combination, the best performance belongs to decision trees with QUEST algorithm (20 variables in women). Generally, data mining can be used in epidemiological studies for different purposes; the decision tree methods determine non-linear relationships among variables by creating a tree structure and they can help to identify risk factors in certain subgroups by creating a threshold or decision boundary. This research employs different data such as clinical information