Mean shift outlier models by M-estimation method and parameter estimation with conic programming

This thesis aims the constitution of Mean Shift Outlier Model (MSOM) on a data set contaminated with outliers, after detection of outliers to not disregard the information possessed by the outliers. The parameters of this model were estimated by using M-estimation method which is a very important and indirect robust outlier detection method. Robustness of M-estimators were combined with the efficiency of Tikhonov Regularization and Least Absolute Shrinkage and Selection Operator (LASSO) to overcome the problem of outliers in linear regression model. Therefore, firstly Tikhonov Regularization and LASSO problems were applied to the Mean Shift Outlier Model (MSOM) based on Huber type M-estimation method. Then, the conic quadratic programming method that uses the interior point method was proposed for solving this problem. Here, The aim is to protect the model from the negative effects of outliers. Moreover, a proposal was introduced on how the calculation of the optimal tuning constant for the given problem. Then, the established models were applied to the data set which was obtained by 21-day operation of a plant for the oxidation of ammonia to nitric acid. For that application MATLAB and MOSEK software packages were used.