V0.2.0                    

Available at

SOL is an open-source library for large-scale sparse online learning, which consists of a family of efficient and scalable sparse online learning algorithms for large-scale online classification tasks. We have offered easy-to-use command-line tools and examples for users and developers. We also have made comprehensive documents available for both beginners and advanced users. SOL is not only a machine learning tool, but also a comprehensive experimental platform for conducting large scale sparse online learning research.

Specifically, SOL consists of a family of first order sparse online learning algorithms as follows

• STG: sparse online learning via truncated graidient (Langford et al., 2009);
• FOBOS: Forward backward splitting (Duchi et al., 2009);
• RDA: Regularized dual averaging(Xiao, 2010);
• RDA_E: Regularized dual averaging(Xiao, 2010); ,a family of second order sparse online learning algorithms as follows
• Ada-FOBOS : adaptive FOBOS (Duchi et al., 2011);
• Ada-RDA: adaptive RDA(Duchi et al., 2011);
• AROW: adaptive regularization of weight vectors (Crammer et al., 2009);
• AROW-TG: adaptive regularization of weight vectors with trunated gradient
• AROW-DA: adaptive regularization of weight vectors with dual averaging and a family of online feature selection algorithms:
• PET: Perceptron with truncation
• FOFS: first order online feature selection (Jialei et al. 2012)
• SOFS: second online feature selection with adaptive regularization of weight vectors for

This document briefly explains the usage of SOL. A more detailed manual can be found from tutorail of SOL.

To get started, please read the ``Quick Start'' section first.

• Structure of folders
• Installation
• Quick Start
• Additional Information

---data example datasets ---doc documentation of the library ---exp_sol python and matlab scripts for l1-regularized sparse online learning experiments, including cross validation and performance evaluation ---exp_ofs python and matlab scripts for online feature selection experiments, including cross validation and performance evaluation ---src source code of the library ---test est code and example use of the library ---tools python scripts to pre-process datasets ---CMakeLists.txt

SOL features a very simple installation procedure. The project is managed by Cmake. There exists a CMakeLists.txt in the root dir of SOL.

1. For linux users

   1. cd to the directory of SOL

   2. make a folder for building the project, like mkdir build

   3. cd to the folder above and call cmake ..

   4. make and you will get an executable SOL in the bin folder

   5. make install and the executable will be copied to the root dir of SOL

2. For windows users

   1. make a folder for building the project

   2. call cmake. Remember to specify the visual studio. For example, if you are using Visual Studio 2012, you can generate the project by

       cmake -G "Visual Studio 11" ..
      

   3. Open the project, Rebuild the ALL_BUILD project and then build the INSTALL project

3. For linux users with Eclipse

   1. cd to the directory of SOL

   2. make a folder for building the project, like mkdir build

   3. cd to the folder above and generate the project by

       cmake -G"Eclipse CDT4 - Unix Makefiles" -D CMAKE_BUILD_TYPE=Debug -DCMAKE_INSTALL_PREFIX=/usr/local ..
      

   4. Install cmakeed plugin into Eclipse.

        cmakeed: http://cmakeed.sourceforge.net/updates/
      

   5. Import the existing project from build into Eclipse.

   6. Build the project throught right click on the project, and then select 'Make Targets' -> 'Build'.

Running SOL without any arguments or with '--help' will produce a message which briefly explains each argument. Below arguments are grouped according to their function.

We provide an example to show how to use SOL and explain the details of how SOL works. The dataset we use will be a6a. Note that only LibSVM datasets are supported by default.

The command for training wit default algorithm is as the following shows.

./SOL -i a6a -m SGD

In this part, we will explain how to induce sparsity of the weight vector and how to tune parameters of algorithms. In STG, we can induce sparsity by:

./SOL -i a6a -opt STG -eta 0.01 -l1 1e-3

Also,we can change the number of steps to truncate the gradients (default is 10).

./SOL -i a6a -opt STG -eta 0.01 -l1 1e-3 -k 1

For details, please check tutorial of SOL.

For any questions and comments, please send your email to chhoi@ntu.edu.sg

Released date: 1 March, 2014.