A few feature requests

User 1251 | 2/10/2015, 3:44:19 PM

1) Allow for a holdout group when using the Vowpal Wabbit wrapper. Right now GraphLab calls VW with the --holdout_off option without any way to to do otherwise.

2) Provide a method for feature extraction from the BoostedTreesClassifier that returns leaf indices similar to the extract_features() method of NeuralNetClassifier or the apply() method of sklearn's RandomForestClassifier implementation.

3) Provide a function that does feature hashing online, rather than handle categorical variables in one large slow initial process as GraphLab currently does.

Comments

User 92 | 2/13/2015, 1:09:23 AM

Hi sbgaucho,

To answer your questions:

1) We currently do not support holdout group for VW. We will consider that for future releases. BTW, All GraphLab Create built-in toolkits support holdout options, and with the coming feature (yes, coming really really soon), you can do a lot more feature engineering like Quadratic Features, Feature Hashing, OneHotEncoder, etc. So try use built-in GraphLab Create classifier and let us know what's missing.

2) This is a very good request and we will consider that.

3) We will have a lot of feature engineering functionalities in coming release. Stay tuned for the announcement!

Ping