SOLVED LING572 Hw8: SVM

Q1 (15 points): Run libSVM on a binary classification task. (a): The data are under hw8/examples/: • train.txt and test.txt are the training and test data in the Mallet format. • train and test are the data in the libSVM format. • You only need to use train and test for this hw. (b): Run libSVM with train as training data, test as test data, and the settings specified in the 2nd-5th columns of Table 1. Fill out the 6th-8th columns of Table 1. Save the model under model.id, where id is the expt id, specified in the first column. (c): You only need to submit model.1 and model.4. Table 1: Results on the binary task Expt id Kernel gamma coef0 degree total sv Training Test Test Acc Acc Acc from Q2 1 linear - - - 2 polynomial 1 0 2 3 polynomial 0.1 0.5 2 4 RBF 0.5 - - 5 sigmoid 0.5 -0.2 - Q2 (60 points): Write an SVM decoder, svm classify.sh, that uses an SVM model created by libSVM to classify test instances. • The command line is: svm classify.sh test data model file sys output • The classifier should be able to handle the four types of kernels specified in Table 1. That is, it should be able to read the kernel type and parameters from the model file and calculate the kernel function accordingly. • test data is in the libSVM data format (e.g., test). • model file is in the libSVM model format (e.g., model ex). The model file stores αiyi for each support vector and ρ (See slide #12-14 in class15 libSVM.pdf). • Each line in sys output (e.g., sys ex) has the format “trueLabel sysLabel fx”: trueLabel is the label in the gold standard, sysLabel is the label produced by the SVM classifier, fx is the value of f(x) = wx − ρ = P i αiyiK(xi , x) − ρ. 1 If f(x) >= 0, then sysLabel should be 0; else sysLabel should be 1. This is different from the convention used in SVM papers/chapters. For other differences between the two conventions, see slide #14 in class15 libSVM.pdf. • Use the model file created in Q1 and test as the test data. Fill out the last column of Table 1. Save the sys output file as sys.id, where id is the expt id in the first column of Table 1. • You only need to submit sys.1 and sys.4. Submission: Submit the following to Canvas: • Your note file readme.(txt | pdf ) that includes Table 1, and any notes that you want the TA to read. • hw.tar.gz that includes all the files specified in dropbox/18-19/572/hw8/submit-file-list, plus any source code (and binary code) used by the shell scripts. • Make sure that you run check hw8.sh before submitting your hw.tar.gz. 2