Indexing Initiative: MTI ML Release Notes

Release Notes for latest version of the MTI ML package for May 1, 2012.

Classifier's score is shown during annotation time using the OVAAnnotatorConfidence class. The confidence can be used to fine tune precision/recall values. The confidence value is not normalized. Future implementations might incorporate this feature.
New learning algorithms based on stochastic gradient descent and loss functions (Hinge loss or modified Huber loss) have been added to the system.
Data set filters allow performing feature selection and sampling. These filters can be placed in a pipeline in the configuration file. Feature selection based on information gain has been added.
Classifiers can be stored on existing classifier files. This allows adding trained models. If a classifier for a given category already exists, it will be overwritten with the new one.