All experiments on (W)SVMs are conducted with LIBSVM-3.18 (Chang & Lin, 2011). The LIBSVM is interfaced with Matlab. It is suggested that the user to download LIBSVM from http://www.csie.ntu.edu.tw/ cjlin/libsvm before the implementation of this toolbox.
The input parameters for the multilevel algorithm are the number of neighbors for coarsening (KNN), the number of neighbors for uncoarsening (KNN_UNCOARSE), the threshold for clustering (Q_dt = U_trainsize), the size of clustering datasets (Param_bag), the threshold for the size of training data for model selection (MoS_UB), the maximum size of the coarsest level (Upperlim), the threshold for the size of coarsening the minority class (Imb_size), the percentage of the nearest clusters for each cluster (per_num), the number of iterations for getting support vectors in clustering (M_LOOP_clusters), Multilevel = 1 if you use multilevel algorithm, else Multilevel=0, Model_Selec = 1 if you use model selection else Model_Selec = 0, resu.Have_test = 0 if we have testdata and traindata in separate files else Have_test = 0 (in this case, we choose 90% for training and 10% for testing)

All input parameters should be given in Multilevel.m for both binary classification. The Composit algorithm is selected as a default algorithm for FLANN.
