CPSC 533: Artificial Intelligence (Fall 1997)

The WEKA machine learning workbench

1. Introduction

This tutorial describes how to use the Waikato Environment for Knowledge Analysis-WEKA. WEKA is a software workbench that provides a consistent graphical user interface to a variety of machine learning algorithms, and facilities of displaying and evaluating the output of those algorithms.

These are some important points you should know:

• WEKA is a prototype and as such is a little rough around the edges.
• WEKA only runs under UNIX.
• WEKA must be used on an X-terminal-preferably a color X-term.

2. Getting started

• WEKA can be run on a Sparc-5 workstation, or csa, csb.

• If you run WEKA on csa or csb, first execute
• setenv DISPLAY xxx:0.0
• where xxx is the name of your workstation.

• Put into your PATH environment variable the directory:

• Start up WEKA:

3. The WEKA main screen

The WEKA main screen contains the following information:

• The dataset(s) being used, including the size of the dataset in features and instances
• The attributes in the dataset and whether they are included or excluded from use
• Data type information about a particular attribute
• The currently selected machine learning scheme
• WEKA processing status and system log information

It also includes menus for selecting machine learning schemes, datasets, activating data and result viewers, generation of dataset summary information and a link to the WEKA WWW home page.

4. A simple example-golf

Here is an example which gets you to run a machine learning algorithm on a simple dataset. The "golf" dataset contains four features that describe the weather and a class that says whether or not golf was played. You are to use the C4.5 scheme to induce a decision tree to build a model describing the decisions to play golf.

In this example you will be working with the dataset:

golf.arff

• Choose the "Open Training Dataset…" option from the "Train" menu and then select the golf dataset.

• Select the machine learning scheme "C4.5" from the "Scheme" menu.
• Click on the "Run C4.5" button to activate the scheme.
• Choose the class attribute-in this case "class".
• Configure the C4.5 algorithm-in this case just click the "Okay" button.

• When the scheme has finished, select the entry off the "Results" menu.

• If you want to you can select the "Unpruned Tree" or "Pruned Tree" option from the "View" menu in the results viewer and look at a graphical representation of the tree.

5. Rules instead of trees-iris

Sometimes it is better to have a rule set than a decision tree. C4.5 allows you to generate a rule set from a decision tree. Use the dataset:

iris.arff

When you come to configuring the C4.5 parameters make sure that the "Rules" check box is checked. You can leave the "Tree" check box checked if you want.

C4.5 generates rules that are evaluated in a top-down fashion and also allows a confusion matrix to be generated.

6. Training and testing-labor negotiations

Sometimes you wish to generate a rule set from some data, and then evaluate how well those rules do on some other data. The first dataset is called the training set and the second dataset the test set.

• Open the dataset: labor-neg-data.arff
• From the "Test" menu select "Open Testing Dataset …" and choose the dataset: labor-neg-test.arff
• Run C4.5 for both trees and rules and look at the results.

7. Other schemes-FOIL and 1R

There are other schemes in WEKA and here two other than C4.5 will be discussed. FOIL is a first-order inductive learner. It generates rules. 1R only generates one simple rule, i.e. the most informative rule.

• Open the dataset: zoo.arff
• Exclude the "animal" attribute by deselecting it.
• Run FOIL with the class being "Type".
• Then run 1R (including 1Rw, 1R, 1R*) on the data.

8. External evaluation-Preval.

Most schemes don't have a built-in evaluators to run test data through. However WEKA has a program called "Preval" (PROLOG Rule EVALuator) that will take a rule set in PROLOG and run a test set again it. If you check the "External Evaluation" option in a scheme's configuration dialog, then the rules will be automatically translated into PROLOG and "Preval" invoked.

Run the labor negotiations data from section 6 through C4.5, FOIL, and other schemes with external evaluation turned on.

9. Extras

Run Netscape and look at the WEKA/Machine Learning home page:

http://www.cs.waikato.ac.nz/~ml