Research On Pattern Classifiers Methods Based On Multi-kernel Learning

Machine learning applications and pattern recognition have gained great attention recently because of the variety of applications based on machine learning techniques,these techniques could make many processes easier and also reduce the amount of human interference(more automation),this thesis discuses classification using Support Vector Machines(SVM),which is an important tool in the field of pattern recognition.SVM in its original form is a linear classifier that can be extended to nonlinearity by using kernels' tricks.In this thesis,the concept of Classification using Kernel methods is discussed with comparing single kernel performance to Linear Combination of kernels,to achieve this goal a number of real benchmark datasets used like Breast Cancer,(ECG)signals information and(EEG)signals,these data sets represent different kinds of features for classification problem,some datasets are for two-classes classification(Binary Classification)and others are for multi-class classification,then the effect of Features normalization is introduced for all the classification datasets under single kernel operations and multi-kernel operations,which shows the different in results between different kernels and different features' formulations.The effect of feature scaling and Normalization on the classification performance will be tested through the experiments.Accuracy,Receiver Operating Characteristic(ROC curve)and Area Under ROC curve(AUC)are used to measure performance.Then the research is extended to Remote Sensing by applying MKL techniques and polynomial kernel to classify satellite images.The main contents of this thesis are as follows:(1)Mainly four kernels were chosen to be used with SVM methods,these kernels are:(Linear,Gaussian,Polynomial and Sigmoid),then eleven benchmark datasets were chosen to apply these kernels under different conditions of normalization and following the best parameters for each kernel to obtain the best performance for each kernel alone.To measure the performance,data sets were divided into training,cross validation and testing sets.(2)Then the previous operations were repeated for the same sets under the same conditions for MKL algorithm,where a new set of parameters were introduced and used to optimize the performance into the best stage that could achieved by the eleven benchmark data sets used in the thesis.then all results(Accuracies and AUCs)were gathered and represented in a proper histograms to show performances achieved by different methods.(3)MKL algorithm and polynomial kernel were applied in remote Sensing application,and the outcome performance was compared to other state of art results from another research for the same application.Finally,the results and experiences obtained from the research experiments werepresented through the discussion and conclusion parts.