Decision Trees Reading: Textbook, “Learning From Examples”, Section 3.

Decision Trees

Reading: Textbook, “Learning From Examples”, Section 3

Day Outlook Temp Humidity Wind PlayTennis

D1 Sunny Hot High Weak NoD2 Sunny Hot High Strong NoD3 Overcast Hot High Weak YesD4 Rain Mild High Weak YesD5 Rain Cool Normal Weak YesD6 Rain Cool Normal Strong NoD7 Overcast Cool Normal Strong YesD8 Sunny Mild High Weak NoD9 Sunny Cool Normal Weak YesD10 Rain Mild Normal Weak YesD11 Sunny Mild Normal Strong YesD12 Overcast Mild High Strong YesD13 Overcast Hot Normal Weak YesD14 Rain Mild High Strong No

Training data:

Decision Trees

• Target concept: “Good days to play tennis”

• Example: <Outlook = Sunny, Temperature = Hot, Humidity =

High, Wind = Strong>

Classification?

• How can good decision trees be automatically constructed?

• Would it be possible to use a “generate-and-test” strategy to find a correct decision tree?

– I.e., systematically generate all possible decision trees, in order of size, until a correct one is generated.

• Why should we care about finding the simplest (i.e., smallest) correct decision tree?

Decision Tree Induction

• Goal is, given set of training examples, construct decision tree that will classify those training examples correctly (and, hopefully, generalize)

• Original idea of decision trees developed in 1960s by psychologists Hunt, Marin, and Stone, as model of human concept learning. (CLS = “Concept Learning System”)

• In 1970s, AI researcher Ross Quinlan used this idea for AI concept learning: – ID3 (“Itemized Dichotomizer 3”), 1979

The Basic Decision Tree Learning Algorithm(ID3)

1. Determine which attribute is, by itself, the most useful one for distinguishing the two classes over all the training data. Put it at the root of the tree.

Outlook



2. Create branches from the root node for each possible value of this attribute. Sort training examples to the appropriate value.

Outlook

Sunny Overcast Rain

D1, D2, D8D9, D11

D3, D7, D12D13

D4, D5, D6D10, D14




3. At each descendant node, determine which attribute is, by itself, the most useful one for distinguishing the two classes for the corresponding training data. Put that attribute at that node.

Outlook

Sunny Overcast Rain

Humidity

Yes

Wind





4. Go to 2, but for the current node.

Note: This is greedy search with no backtracking





4. Go to 2, but for the current node.

Note: This is greedy search with no backtracking

How to determine which attribute is the best classifier for a set of training examples?

E.g., why was Outlook chosen to be the root of the tree?

“Impurity” of a split

• Task: classify as Female or Male

• Instances: Jane, Mary, Alice, Bob, Allen, Doug

• Each instance has two binary attributes: “wears lipstick” and “has long hair”

“Impurity” of a split

Wears lipstick

T F

Jane, Mary, Alice

Pure split Impure split

Bob, Allen, Doug

T F

Jane, Mary, Bob Alice, Allen, Doug

Has long hair

For the each node of the tree we want to choose attribute that gives purest split.

But how to measure degree of impurity of a split ?

Entropy• Let S be a set of training examples.

p+ = proportion of positive examples. p− = proportion of negative examples

• Entropy measures the degree of uniformity or non-uniformity in a collection.

• Roughly measures how predictable collection is, only on basis of distribution of + and − examples.

Entropy

• When is entropy zero?

• When is entropy maximum, and what is its value?

• Entropy gives minimum number of bits of information needed to encode the classification of an arbitrary member of S.

– If p+ = 1, don’t need any bits (entropy 0)

– If p+ = .5, need one bit (+ or -)

– If p+ = .8, can encode collection of {+,-} values using on average less than 1 bit per value

• Can you explain how we might do this?

Entropy of each branch?

Wears lipstick

T F

Jane, Mary, Alice

Pure split Impure split

Bob, Allen, Doug

T F

Jane, Mary, Bob Alice, Allen, Doug

Has long hair



What is the entropy of the “Play Tennis” training set?

• Suppose you’re now given a new example. In absence of any additional information, what classification should you guess?

What is the average entropy of the “Humidity” attribute?

In-class exercise:

• Calculate information gain of the “Outlook” attribute.

Formal definition of Information Gain



Operation of ID3

1. Compute information gain for each attribute.

Outlook

Temperature

Humidity

Wind

ID3’s Inductive Bias

• Given a set of training examples, there are typically many decision trees consistent with that set.

– E.g., what would be another decision tree consistent with the example training data?

• Of all these, which one does ID3 construct?

– First acceptable tree found in greedy search

ID3’s Inductive Bias, continued

• Algorithm does two things:

– Favors shorter trees over longer ones

– Places attributes with highest information gain closest to root.

• What would be an algorithm that explicitly constructs the shortest possible tree consistent with the training data?

ID3’s Inductive Bias, continued

• ID3: Efficient approximation to “find shortest tree” method

• Why is this a good thing to do?

Overfitting

• ID3 grows each branch of the tree just deeply enough to perfectly classify the training examples.

• What if number of training examples is small?

• What if there is noise in the data?

• Both can lead to overfitting– First case can produce incomplete tree– Second case can produce too-complicated tree.

But...what is bad about over-complex trees?

Overfitting, continued

• Formal definition of overfitting:

– Given a hypothesis space H, a hypothesis h H is said to overfit the training data if there exists some alternative h’ H, such that

TrainingError(h) < TrainingError(h’),

but

TestError(h’) < TestError(h).

Overfitting, continuedA

ccur

acy

Size of tree (number of nodes)

test data

training data

Medical data set

Overfitting, continued

• How to avoid overfitting:

– Stop growing the tree early, before it reaches point of perfect classification of training data.

– Allow tree to overfit the data, but then prune the tree.

Pruning a Decision Tree

• Pruning: – Remove subtree below a decision node.

– Create a leaf node there, and assign most common classification of the training examples affiliated with that node.

– Helps reduce overfitting


D1 Sunny Hot High Weak NoD2 Sunny Hot High Strong NoD3 Overcast Hot High Weak YesD4 Rain Mild High Weak YesD5 Rain Cool Normal Weak YesD6 Rain Cool Normal Strong NoD7 Overcast Cool Normal Strong YesD8 Sunny Mild High Weak NoD9 Sunny Cool Normal Weak YesD10 Rain Mild Normal Weak YesD11 Sunny Mild Normal Strong YesD12 Overcast Mild High Strong YesD13 Overcast Hot Normal Weak YesD14 Rain Mild High Strong NoD15 Sunny Hot Normal Strong No

Training data:

Example Outlook

Sunny Overcast Rain

Humidity

Yes

Wind

High Normal

Strong Weak

Temperature

Yes

Hot Mild Cool

No

Yes

Yes

No

No

Example Outlook

Sunny Overcast Rain

Humidity

Yes

Wind

High Normal

Strong Weak

Temperature

Yes

Hot Mild Cool

No

Yes

Yes

No

No

Example Outlook

Sunny Overcast Rain

Humidity

Yes

Wind

High Normal

Strong Weak

Temperature

Yes

Hot Mild Cool

No

Yes

Yes

No

No D9 Sunny Cool Normal Weak YesD11 Sunny Mild Normal Strong YesD15 Sunny Hot Normal Strong No

Example Outlook

Sunny Overcast Rain

Humidity

Yes

Wind

High Normal

Strong Weak

Temperature

Yes

Hot Mild Cool

No

Yes

Yes

No

No D9 Sunny Cool Normal Weak YesD11 Sunny Mild Normal Strong YesD15 Sunny Hot Normal Strong No

Majority: Yes

Example Outlook

Sunny Overcast Rain

Humidity

Yes

Wind

High Normal

Strong Weak

YesNo

No Yes

How to decide which subtrees to prune?

How to decide which subtrees to prune?

Need to divide data into:Training setPruning (validation) setTest set

• Reduced Error Pruning:

– Consider each decision node as candidate for pruning.

– For each node, try pruning node. Measure accuracy of pruned tree over pruning set.

– Select single-node pruning that yields best increase in accuracy over pruning set.

– If no increase, select one of the single-node prunings that does not decrease accuracy.

– If all prunings decrease accuracy, then don’t prune. Otherwise, continue this process until further pruning is harmful.

Simple validation

• Split training data into training set and validation set.

• Use training set to train model with a given set of parameters (e.g., # training epochs). Then use validation set to predict generalization accuracy.

• Finally, use separate test set to test final classifier.

validation

training

training timeor nodes prunedor...

Error rate

stop training/pruning/... here

Miscellaneous

• If you weren’t here last time, see me during the break

• Graduate students (545) sign up for paper presentations– This is optional for undergrads (445)– Two volunteers for Wednesday April 17

• Coursepack on reserve in library

• Course mailing list: [email protected]

Today

• Decision trees

• ID3 algorithm for constructing decision trees

• Calculating information gain

• Overfitting

• Reduced error pruning

• pruning

• Continuous attribute values

• Gain ratio

• UCI ML Repository

• Optdigits data set

• C4.5

• Evaluating classifiers

• Homework 1

Recap from last time



Exercise: What is information gain of Wind?

E(S) = .94

Continuous valued attributes

• Original decision trees: Two discrete aspects:

– Target class (e.g., “PlayTennis”) has discrete values

– Attributes (e.g., “Temperature”) have discrete values

• How to incorporate continuous-valued decision attributes? – E.g., Temperature [0,100]

Continuous valued attributes, continued

• Create new attributes, e.g., Temperaturec true if Temperature >= c, false otherwise.

• How to choose c? – Find c that maximizes information gain.


D1 Sunny 85 High Weak NoD2 Sunny 72 High Strong NoD3 Overcast 62 High Weak YesD4 Rain 60 High Weak YesD5 Rain 20 Normal Strong NoD6 Rain 10 Normal Weak Yes

Training data:

• Sort examples according to values of Temperature found in training setTemperature: 10 20 60 62 72 85PlayTennis: Yes No Yes Yes No No

• Find adjacent examples that differ in target classification.

• Choose candidate c as midpoint of the corresponding interval. – Can show that optimal c must always lie at such a

boundary.

• Then calculate information gain for each candidate c.

• Choose best one.

• Put new attribute Temperaturec in pool of attributes.

Example

Temperature: 10 20 60 62 72 85PlayTennis: Yes No Yes Yes No No

Example


Example


c =15 c =40 c =67

Example


c =15 c =40 c =67

Define new attribute: Temperature15 , with

Values(Temperature15) = { <15 , >=15}


D1 Sunny >=15 High Weak NoD2 Sunny >=15 High Strong NoD3 Overcast >=15 High Weak YesD4 Rain >=15 High Weak YesD5 Rain >=15 Normal Strong NoD6 Rain <15 Normal Weak Yes

Training data:

What is Gain(S, Temperature15)?

• All nodes in decision tree are of the form

Ai

Threshold < Threshold

Alternative measures for selecting attributes

• Recall intuition behind information gain measure:– We want to choose attribute that does the most work in

classifying the training examples by itself.

– So measure how much information is gained (or how much entropy decreased) if that attribute is known.

• However, information gain measure favors attributes with many values.

• Extreme example: Suppose that we add attribute “Date” to each training example. Each training example has a different date.

Day Date Outlook Temp Humidity Wind PlayTennis

D1 3/1 Sunny Hot High Weak NoD2 3/2 Sunny Hot High Strong NoD3 3/3 Overcast Hot High Weak YesD4 3/4 Rain Mild High Weak YesD5 3/5 Rain Cool Normal Weak YesD6 3/6 Rain Cool Normal Strong NoD7 3/7 Overcast Cool Normal Strong YesD8 3/8 Sunny Mild High Weak NoD9 3/9 Sunny Cool Normal Weak YesD10 3/10 Rain Mild Normal Weak YesD11 3/11 Sunny Mild Normal Strong YesD12 3/12 Overcast Mild High Strong YesD13 3/13 Overcast Hot Normal Weak YesD14 3/14 Rain Mild High Strong No

Gain (S, Outlook) = .94 - .694 = .246

What is Gain (S, Date)?

• Date will be chosen as root of the tree.

• But of course the resulting tree will not generalize

Gain Ratio• Quinlan proposed another method of selecting attributes, called “gain

ratio”:

Suppose attribute A splits the training data S into m subsets. Call the subsets S1, S2, ..., Sm.

We can define a set:

The Penalty Term is the entropy of this set.

For example: What is the Penalty Term for the “Date” attribute? How about for “Outlook”?

Day Date Outlook Temp Humidity Wind PlayTennis

D1 3/1 Sunny Hot High Weak NoD2 3/2 Sunny Hot High Strong NoD3 3/3 Overcast Hot High Weak YesD4 3/4 Rain Mild High Weak YesD5 3/5 Rain Cool Normal Weak YesD6 3/6 Rain Cool Normal Strong NoD7 3/7 Overcast Cool Normal Strong YesD8 3/8 Sunny Mild High Weak NoD9 3/9 Sunny Cool Normal Weak YesD10 3/10 Rain Mild Normal Weak YesD11 3/11 Sunny Mild Normal Strong YesD12 3/12 Overcast Mild High Strong YesD13 3/13 Overcast Hot Normal Weak YesD14 3/14 Rain Mild High Strong No

UCI ML Repository

http://archive.ics.uci.edu/ml/

http://archive.ics.uci.edu/ml/datasets/Optical+Recognition+of+Handwritten+Digits

optdigits-picturesoptdigits.infooptdigits.names






Homework 1

• How to download homework and data

• Demo of C4.5

• Accounts on Linuxlab?

• How to get to Linux Lab

• Need help on Linux?

• Newer version C5.0: http://www.rulequest.com/see5-info.html

Decision Trees Reading: Textbook, “Learning From Examples”, Section 3.

Documents