Nayyar A.Zaidi - 2012-05-28

What is RAW?

RAW is a Weka add-on for doing weighting in naive Bayes framework. The weights are learned by optimizing either Mean Square Error or Conditional Log Likelihood.

Where can I get more details about RAW method?

Refer to our paper 'Relaxing the Naive Bayes Attribute Independence Assumption by Attribute Weighting'. The paper is currently under review.

What was RAW's testing environment?

RAW was tested on Weka3-6-7 with Eclipse on Ubuntu environment.

Do I need external Libraries to run RAW?

Yes, RAW uses LBFGS solver to find the function minima. Therefore, you need to configure LBFGS solver before running RAW. For instructions on setting up LBFGS, refer to

<a href="https://github.com/mkobos/lbfgsb_wrapper">LBFGS - Code and Instructions</a>

Which java file runs RAW?

Uncompress the zipped file name weka.jar. Run RAWNB.java in src->main->java->weka->classifiers->bayes->RAW folder

Q. What should I do after downloading?

A. 1) Unzip RAW.zip and place code in weka framework.
2) Configure LBFGS solver, build and make sure it runs properly. Export the folder containing the library.
3) Build weka (using ant) and export weka.jar.
4) Run RAW from command like like >java weka.classifiers.bayes.RAW.RAWNB -t /home/nay/workspace/data/datasets_O/sick.arff -D -O "MSE"

Where will the code go in WEKA framework?

The zip file has the desired directory structure. Currently it goes with in the bayes package in classifiers folder. And DiscreteEstimator file goes in Estimator folder.

You can modify the directories. Just change the package name in the source files.

How many objective functions RAW can optimize?

At the time, two that is MSE and CLL. You can add any objective function based on need as code is fairly modular.

Can I do any regularization?

Yes. Currently there is an option for L2 regularization.

Can you give examples of the input arguments?

-t /home/nayyar/workspace/j_work/weka3-7-5/datasets_O/chess.arff -D
or -t /home/nayyar/workspace/j_work/weka3-7-5/datasets_O/chess.arff -D -O "MSE" -R "None"
etc

What kind of attributes RAW can handle?

Currently, RAW can handle only nominal attributes. Use a -D flag to convert your numeric attributes into discrete.

Is RAW2.0 on the horizon?

Yes, we will soon release RAW2.0. It can handle both numeric and nominal attributes, more regularizations like L1 and beta and ability to weight attribute values instead of attributes.

 

Last edit: Nayyar A.Zaidi 2012-05-28