Artificial Intelligence 11.

Decision Tree Learning

Course V231 Department of Computing Imperial College, London Simon Colton

What to do this Weekend?

If

my parents are visiting

Well go to the cinema

If

not
Then, if its sunny Ill play tennis But if its windy and Im rich, Ill go shopping If its windy and Im poor, Ill go to the cinema If its rainy, Ill stay in

Written as a Decision Tree

Root of tree

Leaves

Using the Decision Tree (No parents on a Sunny Day)

From Decision Trees to Logic

Decision

trees can be written as

Horn clauses in first order logic

Read

from the root to every tip

If this and this and this and this, then do this

In

our example:
If no_parents and sunny_day, then play_tennis no_parents sunny_day play_tennis

Decision Tree Learning Overview

Decision

tree can be seen as rules for performing a categorisation

E.g., what kind of weekend will this be?

Remember

that were learning from examples

Not turning thought processes into decision trees

We

need examples put into categories We also need attributes for the examples
Attributes describe examples (background knowledge) Each attribute takes only a finite set of values

The ID3 Algorithm - Overview

The

major question in decision tree learning

Which nodes to put in which positions Including the root node and the leaf nodes

ID3

uses a measure called Information Gain

Impurity

Based on a notion of entropy

in the data

Used to choose which node to put in next

Node

with the highest information gain is chosen

When there are no choices, a leaf node is put on

Entropy General Idea

From Tom Mitchells book:

In order to define information gain precisely, we begin by defining a measure commonly used in information theory, called entropy that characterizes the (im)purity of an arbitrary collection of examples

Want a notion of impurity in data Imagine a set of boxes and balls in them If all balls are in one box

This is nicely ordered so scores low for entropy Boxes with very few in scores low Boxes with almost all examples in scores low

Calculate entropy by summing over all boxes

Entropy - Formulae
Given

a set of examples, S For examples in a binary categorisation

Where p+ is the proportion of positives And p- is the proportion of negatives

For

examples in categorisations c1 to cn

Where pn is the proportion of examples in cn

Entropy - Explanation

Each category adds to the whole measure When pi is near to 1

(Nearly) all the examples are in this category

So it should score low for its bit of the entropy

log2(pi) gets closer and closer to 0

And this part dominates the overall calculation So the overall calculation comes to nearly 0 (which is good)

When pi is near to 0

(Very) few examples are in this category

So it should score low for its bit of the entropy

log2(pi) gets larger (more negative), but does not dominate

Information Gain
Given

set of examples S and an attribute A

A can take values v1 vm Let Sv = {examples which take value v for attribute A}

Calculate

Gain(S,A)

Estimates the reduction in entropy we get if we know the value of attribute A for the examples in S

An Example Calculation of Information Gain

Suppose

we have a set of examples

S = {s1, s2, s3, s4} In a binary categorisation

With

one positive example and three negative examples The positive example is s1

And

Attribute A

Which takes values v1, v2, v3

S1

takes value v2 for A, S2 takes value v2 for A

S3 takes value v3 for A, S4 takes value v1 for A

First Calculate Entropy(S)

Recall that
Entropy(S) = -p+log2(p+) p-log2(p-)

From binary categorisation, we know that

p+ = and p- =

Hence
Entropy(S) = -(1/4)log2(1/4) (3/4)log2(3/4) = 0.811

Note for users of old calculators:

May need to use the fact that log2(x) = ln(x)/ln(2)

And also note that, by convention:

0*log2(0) is taken to be 0

Calculate Gain for each Value of A

Remember

that

And

that Sv = {set of example with value V for A}

So, Sv1 = {s4}, Sv2 = {s1,s2}, Sv3={s3}

Now,

(|Sv1|/|S|) * Entropy(Sv1) = (1/4) * (-(0/1)*log2(0/1)-(1/1)log2(1/1)) = (1/4) * (0 - (1)log2(1)) = (1/4)(0-0) = 0 Similarly, (|Sv2|/|S|) = 0.5 and (|Sv3|/|S|) = 0

Final Calculation

So,

we add up the three calculations and take them from the overall entropy of S: answer for information gain:

Final

Gain(S,A) = 0.811 (0+1/2+0) = 0.311

The ID3 Algorithm

Given

a set of examples, S

Described by a set of attributes Ai Categorised into categories cj Such that A scores highest for information gain
Relative

1. Choose the root node to be attribute A

to S, i.e., gain(S,A) is the highest over all attributes

2. For each value v that A can take

Draw a branch and label each with corresponding v

Then

see the options in the next slide!

The ID3 Algorithm

For each branch youve just drawn (for value v)

If Sv only contains examples in category c

Then put that category as a leaf node in the tree Then find the default category (which contains the most examples from S)

If Sv is empty

Put this default category as a leaf node in the tree

Otherwise
Remove A from attributes which can be put into nodes Replace S with Sv Find new attribute A scoring best for Gain(S, A) Start again at part 2

Make sure you replace S with Sv

Explanatory Diagram

A Worked Example
Weekend W1 W2 W3 W4 W5 W6 W7 W8 W9 W10 Weather Sunny Sunny Windy Rainy Rainy Rainy Windy Windy Windy Sunny Parents Yes No Yes Yes No Yes No No Yes No Money Rich Rich Rich Poor Rich Poor Poor Rich Rich Rich Decision (Category) Cinema Tennis Cinema Cinema Stay in Cinema Cinema Shopping Cinema Tennis

Information Gain for All of S

S = {W1,W2,,W10} Firstly, we need to calculate:

Entropy(S) = = 1.571 (see notes) For all the attributes we currently have available

Next, we need to calculate information gain

(which is all of them at the moment)

Gain(S, weather) = = 0.7 Gain(S, parents) = = 0.61 Gain(S, money) = = 0.2816 Because this gives us the biggest information gain

Hence, the weather is the first attribute to split on

Top of the Tree

So, this is the top of our tree:

Now, we look at each branch in turn

In particular, we look at the examples with the attribute prescribed by the branch Categorisations are cinema, tennis and tennis for W1,W2 and W10 What does the algorithm say?

Ssunny = {W1,W2,W10}

Set is neither empty, nor a single category

Working with Ssunny

Weekend W1 W2 W10 Weather Sunny Sunny Sunny Parents Yes No No Money Rich Rich Rich Decision Cinema Tennis Tennis

Need to choose a new attribute to split on

Cannot be weather, of course weve already had that Gain(Ssunny, parents) = = 0.918 Gain(Ssunny, money) = = 0

So, calculate information gain again:

Hence we choose to split on parents

Getting to the leaf nodes

If its sunny and the parents have turned up

Then, looking at the table in previous slide

Theres only one answer: go to cinema

If its sunny and the parents havent turned up

Then, again, theres only one answer: play tennis

Hence our decision tree looks like this:

Avoiding Overfitting
Decision

trees can be learned to perfectly fit the data given

This is probably overfitting
The

answer is a memorisation, rather than generalisation

Avoidance

method 1: method 2:

Stop growing the tree before it reaches perfection Grow to perfection, then prune it back aftwerwards
Most

Avoidance

useful of two methods in practice

Appropriate Problems for Decision Tree learning

From

Tom Mitchells book:

Background concepts describe examples in terms of attribute-value pairs, values are always finite in number Concept to be learned (target function)
Has

discrete values

Disjunctive descriptions might be required in the answer

Decision

tree algorithms are fairly robust to errors

In the actual classifications In the attribute-value pairs In missing information