This tool learns a logistic regression model using the [generalized expectation objective] (https://people.cs.umass.edu/~mccallum/papers/druck08sigir.pdf) by Druck, Mann, and McCallum. Usually, logistic regression models are trained by minimizing the cross-entropy objective, which requires labeled instances. Using generalized expectation (GE) objective, we can train logistic regression models by just labeling features. This allows the user to quickly transfer domain knowledge in rapid classifier building, saving labeling effort and alleviating the cold start problem.

• Python (>= 2.7.3)
• Theano (>= 0.8.2)

This tool can be used to train multiclass logistic regression classification model (not just binary model). It comes with both a command-line interface and a Python module interface. Its usage pattern is similar to that of LIBLINEAR.

Learn logistic regression model

python /path/to/ge_cmd.py learn [data] [model] -f [labeled_features]

Each line of the data file is an unlabeled feature vector in sparse format:

[data_id] TAB ([feature_id]:value )+

• data_id: string identifier for the data point.
• feature_id: string identifier for the feature dimension, need not be an integer

Each line of the labeled_features file is a posterior probability distribution of labels upon seeing a feature:

[feature_id] TAB ([label_id]:Pr(label_id|feature_id) )+

• label_id: string identifier for a class label
• Pr(label_id|feature_id): it is OK to provide an estimate of the probability.
• Note: the probability values on each line should add up to 1!

Predict instances using learned model

python /path/to/ge_cmd.py predict [data] [model] [output]

Each line of the output file is in the format:

[data_id] [most_probable_label] ([label_id]:prob )+

Example For a toy example, please take a look at the test/ directory:

cd test/

./test.sh

For more information, please type

python /path/to/ge_cmd.py learn -h

python /path/to/ge_cmd.py predict -h

Please see test_module.py for a preliminary example.

Documentation TBD. Enjoy!