Changes between Version 3 and Version 4 of FAQ
 Timestamp:
 Feb 17, 2009 1:42:14 PM (13 years ago)
Legend:
 Unmodified
 Added
 Removed
 Modified

FAQ
v3 v4 1 Q: Which SVM formulation does SVMQP solve?1 Q: Which SVM formulation does SVMQP and SVMPath solve? 2 2 3 A: SVMQP solvesthe 2norm soft margin SVM classification problem, exactly the same formulation as is solved by3 A: SVMQP and SVMPath solve the 2norm soft margin SVM classification problem, exactly the same formulation as is solved by 4 4 [http://svmlight.joachims.org/ SVMlight]. 5 5 6 6  7 7 8 Q: How can I use SVMQP? 8 Q: What is the difference between SVMQP and SVMPath? 9 10 A: SVMQP is a Fortran implementation of a QP solver to solve one instance of a SVM problem. The implementation contains interior point 11 and active set algorithms. SVMPath is a C++ implementation of the active set method in SVMQP, which is also extended to produce the 12 entire regularization path of solution for a given range of regularization/penalty parameter values. 13 14  15 16 Q: How can I use SVMQP and SVMPath? 9 17 10 18 A: You can compile SVMQP into a library and call it as a subroutine by passing it the data and the labels arranged into 11 appropriate data structures. You also can set the kernel and other parameters for the problem. For details refer to the README file distribued with the source. 19 appropriate data structures. You also can set the kernel and other parameters for the problem. For details refer to the README 20 file distributed with the source. 21 SVMPath can be used as a callable library or as a stand alone code. It read the data from the input file in DOC format  same format as is used 22 by [http://svmlight.joachims.org/ SVMlight]. Various parameters for SVMPath are set in a parameter file. 12 23 13 24  14 25 15 Q: How big are the problems that SVM_QP can handle?26 Q: How big are the problems that SVM_QP/SVMPath can handle? 16 27 17 A: This depends on the version that you use, the available memory and the size of the optim imal active set.28 A: This depends on the version that you use, the available memory and the size of the optimal active set. 18 29 In Linux on and IBM (notsohighend) laptop we were able to solve the '''adult''' and '''web''' problems from the 19 30 [http://www.ics.uci.edu/~mlearn/MLRepository.html UCI repository ] in a matter of minutes or even seconds. However, 20 the number of optimal active support vectors (the examples that are exactly on the margin) did not exceeed 1500 in these tests. If the number of active support vectors is very large and is similar to the number of data points then SVMQP 31 the number of optimal active support vectors (the examples that are exactly on the margin) did not exceed 1500 in these tests. If the number of active support 32 33 vectors is very large and is similar to the number of data points then SVMQP 21 34 will probably be inefficient and will run into memory problems. However, we believe that such cases result in the 22 35 overfitting of the data and, hence, it is questionable whether they should ever be solved. 36 SVMPath has very similar runtime to SVMQP, but may be slightly slower due to C++. Also it may suffer from slowdown when there are too many breakpoints on 37 the regularization path. 23 38 24 39  25 40 26 Q: I would like to try SVMQP , but I don't want to spend too much time setting it up.41 Q: I would like to try SVMQP/SVMPath, but I don't want to spend too much time setting it up. 27 42 28 A: Setting up SVMQP may be easier that is appears from the first glance. Please contact the project manager for help. 29 If you can discuss your specific application we may help you estimate whether SVMQP is the ideal solver for it. 43 A: Setting up may be easier that is appears from the first glance. Please contact the project manager for help. 44 If you can discuss your specific application we may help you estimate whether SVMQP/SVMPath is the ideal solver for it. 45 In the future a Matlab interface is in the plan, please check back. 30 46 47  48 49 50 51 Q: What kind of problem does SINCO solve? 52 53 A: SINCO (Sparse INverse COvariance selection) solves the same problem as is solved by [http://www.princeton.edu/~aspremon/CovSelCode.htm COVSEL] 54 and [http://wwwstat.stanford.edu/~tibs/glasso/index.html Glasso]. In produces a sparse positive definite matrix which is an approximation of the 55 inverse of the covariance matrix of a multivariate Gaussian model. 56 57  58 59 60 61 Q: How can I use SINCO? 62 63 A: SINCO has a Matlab interface (provided) which make it very easy to use in that setting. It can also be used as a callable C++ library. 64 65  66 67 68 69 Q: How good is SINCO? 70 71 A: SINCO is very much under development and testing at the moment. The overall performance is being evaluated. Please check with the project manager if 72 you want to use SINCO.