sofia-ml

Project homepage: http://code.google.com/p/sofia-ml/

==Introduction==

The suite of fast incremental algorithms for machine learning (sofia-ml) can be used for training models for classification or ranking, using several different techniques. This release is intended to aid researchers and practitioners who require fast methods for classification and ranking on large, sparse data sets.

Supported learners include:

* Pegasos SVM
* Stochastic Gradient Descent (SGD) SVM
* Passive-Aggressive Perceptron
* Perceptron with Margins
* ROMMA 

These learners can be configured for classification and ranking, with several sampling methods available.

This implementation gives very fast training times. For example, 100,000 Pegasos SVM training iterations can be performed on data from the CCAT task from the RCV1 benchmark data set (with roughly 780,000 examples) in 0.1 CPU seconds on an ordinary 2.4GHz laptop, with no loss in classification performance compared with other SVM methods. On LETOR learning to rank benchmark tasks, training time with 100,000 Pegasos SVM rank steps complete 0.2 CPU seconds on an ordinary laptop.

The primary computational bottleneck is actually reading the data off of disk; sofia-ml reads and parses data from disk substantially faster than other SVM packages we tested. For example, sofia-ml can read and parse data nearly 10 times faster than the reference Pegasos implementation by Shalev-Shwartz, and nearly 3 times faster than svm_perf by Joachims.

This package provides a commandline utility for training models and using them to predict on new data, and also exposes an API for model training and prediction. The underlying libraries for data sets, weight vectors, and example vectors are also provided for researchers wishing to use these classes to implement other algorithms.

==Quick Start==

These quick-start instructions assume the use of the unix/linux commandline, with g++ installed. There are no external code dependencies.

Step 1 Check out the code:

svn checkout http://sofia-ml.googlecode.com/svn/trunk/sofia-ml sofia-ml-read-only

Step 2 Compile the code:

cd sofia-ml-read-only/src/ make ls ../sofia-ml

Executable should be in main sofia-ml-read-only directory.

If the above did not succeed, run the unit tests to help locate the problem:

make clean make all_test

Step 3 Test the code:

cd .. ./sofia-ml

This should display the set of commandline flags and descriptions.

Train a model on the demo training data.

./sofia-ml --learner_type pegasos --loop_type stochastic --lambda 0.1 --iterations 100000 --dimensionality 150000 --training_file demo/demo.train --model_out demo/model

This should display something like the following:

Reading training data from: demo/demo.train Time to read training data: 0.056134 Time to complete training: 0.075364 Writing model to: demo/model Done.

Test the model on the demo data.

./sofia-ml --model_in demo/model --test_file demo/demo.train --results_file demo/results.txt

Should display the following:

Reading model from: demo/model Done. Reading test data from: demo/demo.train Time to read test data: 0.046729 Time to make test prediction results: 0.000844 Writing test results to: demo/results.txt Done.

Examine a few results in the results file:

head -5 demo/results.txt

Format is: \t. Each line in the results

file corresponds to the same line (in order) in the test file.

1.02114 1 1.18046 1 -1.24609 -1 -1.12822 -1 -1.41046 -1

Note that exact results may vary slightly because these algorithms train

by randomly sampling one example at a time.

Evaluate the results:

perl eval.pl demo/results.txt

Should display something like:

Results for demo/results.txt:

Accuracy 0.9880 (using threshold 0.00) (988/1000) Precision 0.9719 (using threshold 0.00) (311/320) Recall 0.9904 (using threshold 0.00) (311/314) ROC area: 0.999406

Total of 1000 trials.

Note that this evaluation script has limited functionality. For more

options, we recommend using the perf software by Rich Caruana (developed fo

the KDD Cup 2004), available at: http://kodiak.cs.cornell.edu/kddcup/software.html

==Data Format==

This package uses the popular SVM-light sparse data format.