
SVM^{light}Support Vector MachineAuthor: Thorsten Joachims <thorsten@joachims.org> Developed at: Version: 6.01 
SVM^{light} is an implementation of Support Vector Machines (SVMs) in C. The main features of the program are the following:
SVM^{struct}: SVM learning for multivariate and structured outputs like trees, sequences, and sets (available here).
SVM^{perf}: New training algorithm for linear classification SVMs that can be much faster than SVM^{light} for large datasets. It also lets you direcly optimize multivariate performance measures like F1Score, ROCArea, and the Precision/Recall BreakEven Point. (available here).
SVM^{light} is an implementation of Vapnik's Support Vector Machine [Vapnik, 1995] for the problem of pattern recognition, for the problem of regression, and for the problem of learning a ranking function. The optimization algorithms used in SVM^{light} are described in [Joachims, 2002a ]. [Joachims, 1999a]. The algorithm has scalable memory requirements and can handle problems with many thousands of support vectors efficiently.
The software also provides methods for assessing the generalization performance efficiently. It includes two efficient estimation methods for both error rate and precision/recall. XiAlphaestimates [Joachims, 2002a, Joachims, 2000b] can be computed at essentially no computational expense, but they are conservatively biased. Almost unbiased estimates provides leaveoneout testing. SVM^{light} exploits that the results of most leaveoneouts (often more than 99%) are predetermined and need not be computed [Joachims, 2002a].
New in this version is an algorithm for learning ranking functions [Joachims, 2002c]. The goal is to learn a function from preference examples, so that it orders a new set of objects as accurately as possible. Such ranking problems naturally occur in applications like search engines and recommender systems.
Futhermore, this version includes an algorithm for training largescale transductive SVMs. The algorithm proceeds by solving a sequence of optimization problems lowerbounding the solution using a form of local search. A detailed description of the algorithm can be found in [Joachims, 1999c]. A similar transductive learner, which can be thought of as a transductive version of kNearest Neighbor is the Spectral Graph Transducer.
SVM^{light} can also train SVMs with cost models (see [Morik et al., 1999]).
The code has been used on a large range of problems, including text classification [Joachims, 1999c][Joachims, 1998a], image recognition tasks, bioinformatics and medical applications. Many tasks have the property of sparse instance vectors. This implementation makes use of this property which leads to a very compact and efficient representation.
The program is free for scientific use. Please contact me, if you are planning to use the software for commercial purposes. The software must not be further distributed without prior permission of the author. If you use SVM^{light} in your scientific work, please cite as
I would also appreciate, if you sent me (a link to) your papers so that I can learn about your research. The implementation was developed on Solaris 2.5 with gcc, but compiles also on SunOS 3.1.4, Solaris 2.7, Linux, IRIX, Windows NT, and Powermac (after small modifications, see FAQ). The source code is available at the following location:
http://download.joachims.org/svm_light/v6.01/svm_light.tar.gz
If you just want the binaries, you can download them for the following systems:
Please send me email and let me know that you got svmlight. I will put you on my mailing list to inform you about new versions and bugfixes. SVM^{light} comes with a quadratic programming tool for solving small intermediate quadratic programming problems. It is based on the method of Hildreth and D'Espo and solves small quadratic programs very efficiently. Nevertheless, if for some reason you want to use another solver, the new version still comes with an interface to PR_LOQO. The PR_LOQO optimizer was written by A. Smola. It can be requested from http://www.kernelmachines.org/code/prloqo.tar.gz.
To install SVM^{light} you need to download svm_light.tar.gz. Create a new directory:
mkdir svm_light
Move svm_light.tar.gz to this directory and unpack it with
gunzip c svm_light.tar.gz  tar xvf 
Now execute
make or make all
which compiles the system and creates the two executables
If you do not want to use the builtin optimizer but PR_LOQO instead, create a subdirectory in the svm_light directory with
mkdir pr_loqo
and copy the files pr_loqo.c and pr_loqo.h in there. Now execute
make svm_learn_loqo
If the system does not compile properly, check this FAQ.
This section explains how to use the SVM^{light} software. A good introduction to the theory of SVMs is Chris Burges' tutorial.
SVM^{light} consists of a learning module (svm_learn) and a classification module (svm_classify). The classification module can be used to apply the learned model to new examples. See also the examples below for how to use svm_learn and svm_classify.
svm_learn is called with the following parameters:
svm_learn [options] example_file model_file
Available options are:
General options: ?  this help v [0..3]  verbosity level (default 1) Learning options: z {c,r,p}  select between classification (c), regression (r), and preference ranking (p) (see [Joachims, 2002c]) (default classification) c float  C: tradeoff between training error and margin (default [avg. x*x]^1) w [0..]  epsilon width of tube for regression (default 0.1) j float  Cost: costfactor, by which training errors on positive examples outweight errors on negative examples (default 1) (see [Morik et al., 1999]) b [0,1]  use biased hyperplane (i.e. x*w+b0) instead of unbiased hyperplane (i.e. x*w0) (default 1) i [0,1]  remove inconsistent training examples and retrain (default 0) Performance estimation options: x [0,1]  compute leaveoneout estimates (default 0) (see [5]) o ]0..2]  value of rho for XiAlphaestimator and for pruning leaveoneout computation (default 1.0) (see [Joachims, 2002a]) k [0..100]  search depth for extended XiAlphaestimator (default 0) Transduction options (see [Joachims, 1999c], [Joachims, 2002a]): p [0..1]  fraction of unlabeled examples to be classified into the positive class (default is the ratio of positive and negative examples in the training data) Kernel options: t int  type of kernel function: 0: linear (default) 1: polynomial (s a*b+c)^d 2: radial basis function exp(gamma ab^2) 3: sigmoid tanh(s a*b + c) 4: user defined kernel from kernel.h d int  parameter d in polynomial kernel g float  parameter gamma in rbf kernel s float  parameter s in sigmoid/poly kernel r float  parameter c in sigmoid/poly kernel u string  parameter of user defined kernel Optimization options (see [Joachims, 1999a], [Joachims, 2002a]): q [2..]  maximum size of QPsubproblems (default 10) n [2..q]  number of new variables entering the working set in each iteration (default n = q). Set n<q to prevent zigzagging. m [5..]  size of cache for kernel evaluations in MB (default 40) The larger the faster... e float  eps: Allow that error for termination criterion [y [w*x+b]  1] = eps (default 0.001) h [5..]  number of iterations a variable needs to be optimal before considered for shrinking (default 100) f [0,1]  do final optimality check for variables removed by shrinking. Although this test is usually positive, there is no guarantee that the optimum was found if the test is omitted. (default 1) y string > if option is given, reads alphas from file with given and uses them as starting point. (default 'disabled') # int > terminate optimization, if no progress after this number of iterations. (default 100000) Output options: l char  file to write predicted labels of unlabeled examples into after transductive learning a char  write all alphas to this file after learning (in the same order as in the training set)
A more detailed description of the parameters and how they link to the respective algorithms is given in the appendix of [Joachims, 2002a].
The input file example_file contains the training examples. The first lines may contain comments and are ignored if they start with #. Each of the following lines represents one training example and is of the following format:
The target value and each of the feature/value pairs are separated by a space character. Feature/value pairs MUST be ordered by increasing feature number. Features with value zero can be skipped. The string <info> can be used to pass additional information to the kernel (e.g. non feature vector data).
In classification mode, the target value denotes the class of the example. +1 as the target value marks a positive example, 1 a negative example respectively. So, for example, the line
1 1:0.43 3:0.12 9284:0.2 # abcdef
specifies a negative example for which feature number 1 has the value 0.43, feature number 3 has the value 0.12, feature number 9284 has the value 0.2, and all the other features have value 0. In addition, the string abcdef is stored with the vector, which can serve as a way of providing additional information for user defined kernels. A class label of 0 indicates that this example should be classified using transduction. The predictions for the examples classified by transduction are written to the file specified through the l option. The order of the predictions is the same as in the training data.
In regression mode, the <target> contains the realvalued target value.
In ranking mode [Joachims, 2002c], the target value is used to generated pairwise preference constraints (see STRIVER). A preference constraint is included for all pairs of examples in the example_file, for which the target value differs. The special feature "qid" can be used to restrict the generation of constraints. Two examples are considered for a pairwise preference constraint only, if the value of "qid" is the same. For example, given the example_file
3 qid:1 1:0.53 2:0.12
2 qid:1 1:0.13 2:0.1
7 qid:2 1:0.87 2:0.12
a preference constraint is included only for the first and the second example(ie. the first should be ranked higher than the second), but not with the third example, since it has a different "qid".
In all modes, the result of svm_learn is the model which is learned from the training data in example_file. The model is written to model_file. To make predictions on test examples, svm_classify reads this file. svm_classify is called with the following parameters:
svm_classify [options] example_file model_file output_file
Available options are:
h Help. v [0..3] Verbosity level (default 2). f [0,1] 0: old output format of V1.0 1: output the value of decision function (default)
The test examples in example_file are given in the same format as the training examples (possibly with 0 as class label). For all test examples in example_file the predicted values are written to output_file. There is one line per test example in output_file containing the value of the decision function on that example. For classification, the sign of this value determines the predicted class. For regression, it is the predicted value itself, and for ranking the value can be used to order the test examples. The test example file has the same format as the one for svm_learn. Again, <class> can have the value zero indicating unknown.
If you want to find out more, try this FAQ.
You will find an example text classification problem at
http://download.joachims.org/svm_light/examples/example1.tar.gz
Download this file into your svm_light directory and unpack it with
gunzip c example1.tar.gz  tar xvf 
This will create a subdirectory example1. Documents are represented as feature vectors. Each feature corresponds to a word stem (9947 features). The task is to learn which Reuters articles are about "corporate acquisitions". There are 1000 positive and 1000 negative examples in the file train.dat. The file test.dat contains 600 test examples. The feature numbers correspond to the line numbers in the file words. To run the example, execute the commands:
svm_learn example1/train.dat example1/model
svm_classify example1/test.dat example1/model example1/predictions
The accuracy on the test set is printed to stdout.
To try out the transductive learner, you can use the following dataset (see also Spectral Graph Transducer). I compiled it from the same Reuters articles as used in the example for the inductive SVM. The dataset consists of only 10 training examples (5 positive and 5 negative) and the same 600 test examples as above. You find it at
http://download.joachims.org/svm_light/examples/example2.tar.gz
Download this file into your svm_light directory and unpack it with
gunzip c example2.tar.gz  tar xvf 
This will create a subdirectory example2. To run the example, execute the commands:
svm_learn example2/train_transduction.dat example2/model
svm_classify example2/test.dat example2/model example2/predictions
The classification module is called only to get the accuracy printed. The transductive learner is invoced automatically, since train_transduction.dat contains unlabeled examples (i. e. the 600 test examples). You can compare the results to those of the inductive SVM by running:
svm_learn example2/train_induction.dat example2/model
svm_classify example2/test.dat example2/model example2/predictions
The file train_induction.dat contains the same 10 (labeled) training examples as train_transduction.dat.
For the ranking SVM [Joachims, 2002c], I created a toy example. It consists of only 12 training examples in 3 groups and 4 test examples. You find it at
http://download.joachims.org/svm_light/examples/example3.tar.gz
Download this file into your svm_light directory and unpack it with
gunzip c example3.tar.gz  tar xvf 
This will create a subdirectory example3. To run the example, execute the commands:
svm_learn z p example3/train.dat example3/model
svm_classify example3/test.dat example3/model example3/predictions
The output in the predictions file can be used to rank the test examples. If you do so, you will see that it predicts the correct ranking. The values in the predictions file do not have a meaning in an absolute sense. They are only used for ordering.
It can also be interesting to look at the "training error" of the ranking SVM. The equivalent of training error for a ranking SVM is the number of training pairs that are misordered by the learned model. To find those pairs, one can apply the model to the training file:
svm_classify example3/train.dat example3/model example3/predictions.train
Again, the predictions file shows the ordering implied by the model. The model ranks all training examples correctly.
Note that ranks are comparable only between examples with the same qid. Note also that the target value (first value in each line of the data files) is only used to define the order of the examples. Its absolute value does not matter, as long as the ordering relative to the other examples with the same qid remains the same.
If you find bugs or you have problems with the code you cannot solve by yourself, please contact me via email <svmlight@ls8.cs.unidortmund.de.
This software is free only for noncommercial use. It must not be distributed without prior permission of the author. The author is not responsible for implications from the use of this software.
[Joachims, 2002a] 
Thorsten Joachims, Learning to Classify Text Using Support Vector Machines. Dissertation, Kluwer, 2002. 
[Joachims, 2002c] 
T.
Joachims, Optimizing Search Engines Using Clickthrough Data, Proceedings
of the ACM Conference on Knowledge Discovery and Data Mining (KDD), ACM, 2002. Online [Postscript] [PDF] 
[Klinkenberg, Joachims, 2000a] 
R. Klinkenberg and T. Joachims, Detecting Concept Drift with Support Vector Machines. Proceedings of the Seventeenth International Conference on Machine Learning (ICML), Morgan Kaufmann, 2000. 
[Joachims, 2000b] 
T. Joachims, Estimating the Generalization Performance of a SVM Efficiently. Proceedings of the International Conference on Machine Learning, Morgan Kaufman, 2000. 
[Joachims, 1999a] 
T. Joachims, 11 in: Making largeScale SVM Learning Practical. Advances in Kernel Methods  Support Vector Learning, B. Schölkopf and C. Burges and A. Smola (ed.), MIT Press, 1999. 
[Joachims, 1999c] 
Thorsten Joachims, Transductive Inference for Text Classification using Support Vector Machines. International Conference on Machine Learning (ICML), 1999. 
[Morik et al., 1999a] 
K. Morik, P. Brockhausen, and T. Joachims, Combining statistical learning with a knowledgebased approach  A case study in intensive care monitoring. Proc. 16th Int'l Conf. on Machine Learning (ICML99), 1999. 
[Joachims, 1998a] 
T. Joachims, Text Categorization with Support Vector Machines: Learning with Many Relevant Features. Proceedings of the European Conference on Machine Learning, Springer, 1998. 
[Joachims, 1998c] 
Thorsten Joachims, Making LargeScale SVM Learning Practical. LS8Report, 24, Universität Dortmund, LS VIIIReport, 1998. 
[Vapnik, 1995a] 
Vladimir N. Vapnik, The Nature of Statistical Learning Theory. Springer, 1995. 
Last modified July 20th, 2004 by Thorsten Joachims <thorsten@joachims.org>