CS 540 Teaching Assistant Information

HW1 : Decision Trees

• What should the printed tree format look like?

There is no one set format for this. In general, you want to print out the tree with indentation signifying the node's level in the tree. You should also print out the test that is being performed at each node and the classification if it's a leaf node.

Here's an example of a possible format, where "shape", "age", and "color" are example features and the classifications are '+' or '-':

shape = circle:
  age = young:  +
  age = old:
    color = red: - 
    color = green: +
shape = square:
  color = red: +
  color = green: -

It is reasonably easy to write Java code to produce the above using recursion, where a "treeDepth" (an int) is passed in and used to determine the amount of leading spaces per line (say 2 x treeDepth spaces). If you're having trouble with this, come talk to me at my office hours and I can help you out.


• Which files should I turn in for the programming section?

You only need to turn in the Java source files (.java). You should delete the .class files from your directory before you submit. If you are using an IDE for development (Eclipse, NetBeans, etc.) make sure your program can be compiled from the command-line. A good way to test this is to copy the files to one of the lab machines. Put all of the .java files in a directory and type the following commands:

javac *.java
java BuildAndTestDecisionTree (parameters)

This should compile and run your program from the command-line. This is how I will be grading your assignment so it's important that this works correctly. If you have problems with this, no problem, just send me an email or come to my office hours and I can help you.


• How should I break ties in the plurality-value method in my program?

The ID-3 algorithm calls the plurality-value method when it has run out of examples in a leaf node or if there are no more attributes to split the examples on. Normally this method returns the majority label of the parent examples (in the first case) or the current examples (in the second case). However, in the case of a tie, you should follow these steps:

1. Take the majority label of the whole training dataset. We often do this in machine learning because it gives us the best chance of generalization.
2. If that is also tied (in the case of the Voting dataset), pick the label which comes first alphabetically.

Note that we do not want you to use randomness in your decision tree construction. This will be to facilitate grading your programs. In "real life" we often would use randomness to break these ties.