HowTo - Using LIBSVM

Data format [1] training_data / testing_data +1 1:6.240913 2:0.117454 3:-0.957491 4:1.653823 5:0.929491    6:-1.036259 7:-1.025772 8:0.372377 9:0.273566 10:-0.493052    11:-0.583311 12:-0.218060 13:-0.376675 14:-0.589258 15:-0.346050    16:0.125185 17:0.106718 18:-0.285916 19:-0.459112 20:-0.170765  -1 1:10.284770 2:1.785967 3:-2.979066 4:-2.174549 5:-1.434280     6:-1.306482 7:-0.548775 8:0.777055 9:0.275654 10:-1.284672     11:-1.663850 12:-0.432742 13:0.273012 14:-0.137017 15:-0.533320     16:-0.255564 17:-0.193427 18:-0.584661 19:-0.668189 20:-0.118506 -1 1:8.184908 2:0.312909 3:-2.742637 4:-0.217420 5:0.472176     6:-1.277293 7:-2.298117 8:-0.996349 9:-0.141399 10:-0.481425     11:-0.589631 12:0.220194 13:0.382807 14:-0.349594 15:-0.694144     16:-0.155397 17:0.011160 18:-0.555094 19:-0.893452 20:-0.488398 +1 1:8.695444 2:0.383987 3:-1.195099 4:2.320266 5:1.958957     6:-0.578233 7:-1.353031 8:-0.364041 9:-0.357977 10:-0.794986     11:-0.867561 12:-0.755012 13:-1.061063 14:-1.202318 15:-0.848766     16:-0.270570 17:-0.199191 18:-0.510582 19:-0.540136 20:-0.134563 ... --- each line is an instance +1 / -1 --- label  1:6.240913   --- the value of Feature 1 is 6.240913 . For prediction (if you don't know the labels of instances), I just put 0 where +1/-1 is put. It doesn't matter what you put here.  You only need to check the predict result in the predict file (see the "Predict" session below). Scale Get scaling file range, scale training_data into training_data.scaled $ ~/libsvm-3.11$ svm-scale -l -1 -u 1 -s range training_data > training_data.scaled Train Get parameters for C-SVM classification (-c ? -v ?) $ ~/libsvm-3.11$ cd tools $ ~/libsvm-3.11/tools$ python grid.py ../training_data.scaledCross validation (-v) $ ~/libsvm-3.11/tools$ cd .. $ ~/libsvm-3.11$ svm-train -c 2 -g 8 -v 4 training_data.scaled Get classifying model $ ~/libsvm-3.11$ svm-train -c 2 -g 8 training_data.scaled After this, there will be a .model file in the current file folder. Test Scale testing_data $ ~/libsvm-3.11$ svm-scale -r range testing_data > testing_data.scaledPredict $ ~/libsvm-3.11$ svm-predict testing_data.scaled training_data.scaled.model predict Predicted labels will be in the predict file in the current file folder. predict -1 -1 1 -1 . . . Retrieve w and b from model [2] w = model.SVs' * model.sv_coef; % the first number of each line is the coefficient b = -model.rho; if model.Label(1) == -1 w = -w; b = -b; end Reference [1] http://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf [2] http://www.csie.ntu.edu.tw/~cjlin/libsvm/faq.html#f804