Decision Trees

Lecturer: Ji Liu

Thank Jerry Zhu for sharing his slides

[ Some slides from Andrew Moore http://www.cs.cmu.edu/~awm/tutorials and Chuck Dyer, with permission.]
 x
• The input
• These names are the same: example,
point, instance, item, input
• Usually represented by a feature
vector
– These names are the same: attribute,
feature
– For decision trees, we will especially
focus on discrete features (though
continuous features are possible, see
end of slides)
 y
• The output
• These names are the same: label,
target, goal
• It can be
– Continuous, as in our population
predictionRegression
– Discrete, e.g., is this mushroom x edible
or poisonous?  Classification
 Evaluating classifiers
• During training
– Train a classifier from a training set (x1,y1),
(x2,y2), …, (xn,yn).
• During testing
– For new test data xn+1…xn+m, your classifier
generates predicted labels y’n+1… y’n+m
• Test set accuracy:
– You need to know the true test labels yn+1…
yn+m n +m
1
– Test set accuracy: acc = m ∑ 1 yi = y ' i
i=n+ 1
– Test set error rate = 1 – acc
 Decision Trees
• One kind of classifier (supervised
learning)
• Outline:
– The tree
– Algorithm
– Mutual information of questions
– Overfitting and Pruning
– Extensions: real-valued features,
treerules, pro/con
 Akinator: Decision Tree
• http://en.akinator.com/personnages/
 A Decision Tree
• A decision tree has 2 kinds of nodes
1. Each leaf node has a class label,
determined by majority vote of training
examples reaching that leaf.
2. Each internal node is a question on
features. It branches out according to
the answers.
Automobile Miles-per-gallon
prediction
mpg cylinders displacement horsepower weight acceleration modelyear maker

good 4 low low low high 75to78 asia

bad 6 medium medium medium medium 70to74 america
bad 4 medium medium medium low 75to78 europe
bad 8 high high high low 70to74 america
bad 6 medium medium medium medium 70to74 america
bad 4 low medium low medium 70to74 asia
bad 4 low medium low low 70to74 asia
bad 8 high high high low 75to78 america
: : : : : : : :
: : : : : : : :
: : : : : : : :
bad 8 high high high low 70to74 america
good 8 high medium high high 79to83 america
bad 8 high high high low 75to78 america
good 4 low low low low 79to83 america
bad 6 medium medium medium high 75to78 america
good 4 medium low low low 79to83 america
good 4 low low medium high 79to83 america
bad 8 high high high low 70to74 america
good 4 low medium low medium 75to78 europe
bad 5 medium medium medium medium 75to78 europe
A very small decision tree
Internal node
question: “what is the
number of
cylinders”?

Leaves: classify by
majority vote
 A bigger decision tree
question: “what is the
value of
horsepower”?
question: “what is the
value of maker”?

Predict “good” is also reasonable by following its parent node instead of the root node.
The full 1. Do not split when all
examples have the same
decision label

tree

2. Can not split when we

run out of questions
 Decision tree algorithm
buildtree(examples, questions, default)
/* examples: a list of training examples
questions: a set of candidate questions, e.g., “what’s the value
of feature xi?”
default: default label prediction, e.g., over-all majority vote */
IF empty(examples) THEN return(default)
IF (examples have same label y) THEN return(y)
IF empty(questions) THEN return(majority vote in
examples)
q = best_question(examples, questions)
Let there be n answers to q
– Create and return an internal node with n children
– The ith child is built by calling
buildtree({example|q=ith answer}, questions\{q}, default)
 The best question
• What do we want: pure leaf nodes, i.e. all
examples having (almost) the same y.
• A good question  a split that results in
pure child nodes
• How do we measure the degree of purity
induced by a question? Here’s one
possibility (Max-Gain in book):
mutual information
(a.k.a. information gain)
A quantity from information theory
Entropy (Impurity Measure)
• At the current node, there are n=n1+…+nk
examples
– n1 examples have label y1
– n2 examples have label y2
–…
– nk examples have label yk
• What’s the impurity of the node?
• Turn it into a game: if I put these
examples in a bag, and grab one at
random, what is the probability the
example has label yi?
 Entropy (Impurity Measure)
• Probability estimated from samples:
 with probability p1=n1/n the example has label y1
 with probability p2=n2/n the example has label y2
 …
 with probability pk=nk/n the example has label yk
• p1+p2+…+pk=1
• The “outcome” of the draw is a random variable y with
probability (p1, p2, …, pk)
• What’s the impurity of the node  what’s the
uncertainty of y in a random drawing?
 Entropy (Impurity Measure)

• Interpretation: The number of yes/no questions (bits)

needed on average to pin down the value of y in a
random drawing

H(y)= H(y)= H(y)=

Entropy (Impurity Measure)

biased
coin
Jerry’s coin

p(head)=0.5 p(head)=0.51 p(head)=1

p(tail)=0.5 p(tail)=0.49 p(tail)=0
H=1 H=0.9997 H=0 (Why?)
 Excellent Video for Entropy
https://www.youtube.com/watch?
v=R4OlXb9aTvQ

• Entropy roughly measures the average

number of yes/no questions we need to
ask to figure out the class label of an
object without any additional attribute
information.
 Conditional entropy
k
H (Y ∣X =v )=∑ −Pr (Y = y i∣X =v )log 2 Pr (Y = y i∣ X=v )
i=1

H (Y ∣X )= ∑ Pr ( X =v ) H (Y ∣X =v )
v :values of X

• Y: label. X: a question (e.g., a feature). v: an

answer to the question
• Pr(Y|X=v): conditional probability
• H(Y|X) estimates the average number of y/n
questions required after know the attribute
information X
 Information gain
• Information gain, or mutual
information
I (Y ; X )=H (Y )−H (Y ∣X )
• Choose question (feature) X which
maximizes I(Y;X).
 Example
• Features: color, shape, size
• What’s the best question at root?

+ -
 The training set
Example Color Shape Size Class

1 Red Square Big +

2 Blue Square Big +
3 Red Circle Big +
4 Red Circle Small -
5 Green Square Small -
6 Green Square Big -

H(class)=
H(class | color)=
Example Color Shape Size Class

1 Red Square Big +

2 Blue Square Big +
3 Red Circle Big +
4 Red Circle Small -
5 Green Square Small -
6 Green Square Big -

H(class)= H(3/6,3/6) = 1
H(class | color)= 3/6 * H(2/3,1/3) + 1/6 * H(1,0) + 2/6 * H(0,1)

blue is + green is -
2 of the
3 out of 6 1 out of 6
red are + 2 out of 6
are red is blue are green
Example Color Shape Size Class

1 Red Square Big +

2 Blue Square Big +
3 Red Circle Big +
4 Red Circle Small -
5 Green Square Small -
6 Green Square Big -

H(class)= H(3/6,3/6) = 1
H(class | color)= 3/6 * H(2/3,1/3) + 1/6 * H(1,0) + 2/6 * H(0,1)
I(class; color) = H(class) – H(class | color) = 0.54 bits
Example Color Shape Size Class

1 Red Square Big +

2 Blue Square Big +
3 Red Circle Big +
4 Red Circle Small -
5 Green Square Small -
6 Green Square Big -

H(class)= H(3/6,3/6) = 1
H(class | shape)= 4/6 * H(1/2, 1/2) + 2/6 * H(1/2,1/2)
I(class; shape) = H(class) – H(class | shape) = 0 bits

Shape tells us
nothing about
the class!
Example Color Shape Size Class

1 Red Square Big +

2 Blue Square Big +
3 Red Circle Big +
4 Red Circle Small -
5 Green Square Small -
6 Green Square Big -

H(class)= H(3/6,3/6) = 1
H(class | size)= 4/6 * H(3/4, 1/4) + 2/6 * H(0,1)
I(class; size) = H(class) – H(class | size) = 0.46 bits
Example Color Shape Size Class

1 Red Square Big +

2 Blue Square Big +
3 Red Circle Big +
4 Red Circle Small -
5 Green Square Small -
6 Green Square Big -

I(class; color) = H(class) – H(class | color) = 0.54 bits

I(class; shape) = H(class) – H(class | shape) = 0 bits
I(class; size) = H(class) – H(class | size) = 0.46 bits

 We select color as the question at root

 Overfitting
• Overfitting happens if the
prediction model is
overcomplicated while the
training data is few.
• Another perspective to say
overfitting is the model fits the
training data perfectly.
• https://www.youtube.com/watch?v=iILj9g8xObc
Example: Overfitting in SVM

x2

x1
Example: Overfitting in regression:
Predicting US Population
x=Year y=Million
1900 75.995 • We have
1910 91.972 some
1920 105.71 training
1930 123.2
1940 131.67 data
1950 150.7 (n=11)
1960 179.32
1970 203.21
• What will
1980 226.51 the
1990 249.63 population
2000 281.42
be in 2020?
 Regression: Polynomial fit
• The degree d (complexity of the model) is
important
d d −1
f ( x )=c d x +c d −1 x +⋯+c 1 x +c 0
• Fit (=learn) coefficients cd, … c0 to
minimize Mean Squared Error (MSE) on
training data
n
1 2
MSE= ∑ ( y i −f ( x i ) )
n i=1
 Overfitting
• As d increases, MSE on training data
improves, but prediction outside training data
worsens
degree=0 MSE=4181.451643
degree=1 MSE=79.600506
degree=2 MSE=9.346899
degree=3 MSE=9.289570
degree=4 MSE=7.420147
degree=5 MSE=5.310130
degree=6 MSE=2.493168
degree=7 MSE=2.278311
degree=8 MSE=1.257978
degree=9 MSE=0.001433
degree=10 MSE=0.000000
 Overfitting: Toy Example
• Predict if the outcome of throwing a
die is “6” from its (color, size)
• Color = {red, blue}, Size={small,
large}
• Three training samples:
– X1 = (red, large), y1 = not 6
– X2 = (blue, small), y2 = not 6
– X3 = (blue, large), y3 = 6
 Overfitting: Example for
Decision Tree
• Three training samples:
– X1 = (red, large), y1 = not 6
– X2 = (blue, small), y2 = not 6
Root
– X3 = (blue, large), y3 = 6 Color ?
(1, 2)

Red Blue
Size?
(0,1) (1, 1)
Not 6
Large Small

(1, 0) (0, 1)
It is 6 Not 6
 Toy Example
• Assume “color” and “size” are
independent attributes for any die
• Assume P(red)=P(blue)=1/2,
P(large)=P(small)=1/2
• The prediction accuracy for this
decision tree is 1-(1/2*1/6+1/4*5/6 +
1/4*1/6)=2/3
 Toy Example
• If the decision tree only has the root
node, we predict all new instances as
“Not 6”.
• The accuracy is 5/6 > 2/3 Root
(1, 2)
Not 6
 Overfit a decision tree
Output y = copy of
e,
Five inputs, all bits, are
Except a random
generated in all 32
25% of the records
possible combinations
have y set to the
opposite of e

a b c d e y
0 0 0 0 0 0
0 0 0 0 1 0
records

0 0 0 1 0 0
32

0 0 0 1 1 1
0 0 1 0 0 1
: : : : : :
1 1 1 1 1 1
 Overfit a decision tree
• The test set is constructed similarly
– y=e, but 25% the time we corrupt it by
y=¬e
– The corruptions in training and test sets
are independent
• The training and test sets are the same,
except
– Some y’s are corrupted in training, but not
in test
– Some y’s are corrupted in test, but not in
training
 Overfit a decision tree
• We build a full tree on the training
set Root

e=0 e=1

a=0 a=1 a=0 a=1

Training set accuracy = 100%

25% of these training leaf node labels will be corrupted (≠e)
 Overfit a decision tree
• And classify the test data with the tree

Root

e=0 e=1

a=0 a=1 a=0 a=1

25% of the test examples are corrupted – independent of training data

 Overfit a decision tree

On average:
• ¾ training data uncorrupted
– ¾ of these are uncorrupted in test – correct
labels
– ¼ of these are corrupted in test – wrong
• ¼ training data corrupted
– ¾ of these are uncorrupted in test – wrong
– ¼ of these are also corrupted in test – correct
labels
• Test accuracy = ¾ * ¾ + ¼ * ¼ = 5/8 = 62.5%
 Overfit a decision tree
• But if we knew a,b,c,d are irrelevant features and don’t use them in the
tree…

Pretend they don’t

exist

a b c d e y
0 0 0 0 0 0
0 0 0 0 1 0
0 0 0 1 0 0
0 0 0 1 1 1
0 0 1 0 0 1
: : : : : :
1 1 1 1 1 1
 Overfit a decision tree
• The tree would be

Root

e=0 e=1

In training data, about ¾ y’s are In training data, about ¾ y’s are
0 here. Majority vote predicts 1 here. Majority vote predicts
y=0 y=1

In test data, ¼ y’s are different from e.

test accuracy = ?
 Overfit a decision tree
• The tree would be

Root

e=0 e=1

In training data, about ¾ y’s are In training data, about ¾ y’s are
0 here. Majority vote predicts 1 here. Majority vote predicts
y=0 y=1

In test data, ¼ y’s are different from e.

test accuracy = ¾ = 75% (better!)

Full tree test accuracy = ¾ * ¾ + ¼ * ¼ = 5/8 = 62.5%

 Overfit a decision tree
• In the full tree, we overfit by learning non-existent relations (noise)

Root

e=0 e=1

a=0 a=1 a=0 a=1

 Avoid overfitting: pruning
Pruning with a tuning set
1. Randomly split data into TRAIN and
TUNE, say 70% and 30%
2. Build a full tree using only TRAIN
3. Prune the tree down on the TUNE set.
On the next page you’ll see a greedy
version.
 Pruning
Prune(tree T, TUNE set)
1. Compute T’s accuracy on TUNE, call it A(T)
2. For every internal node N in T:
a) New tree TN = copy of T, but prune (delete) the subtree
under N.
b) N becomes a leaf node in TN. The label is the majority
vote of TRAIN examples reaching N.
c) A(TN) = TN’s accuracy on TUNE
3. Let T* be the tree (among the TN’s and T) with
the largest A(). Set TT* /* prune */
4. Repeat from step 1 until no more improvement
available. Return T.
 Real-valued features
• What if some (or all) of the features
x1, x2, …, xk are real-valued?
• Example: x1=height (in inches)
• Idea 1: branch on each possible
numerical value.
 Real-valued features
• What if some (or all) of the features x1, x2, …, xk are real-valued?
• Example: x1=height (in inches)
• Idea 1: branch on each possible numerical value. (fragments the
training data and prone to overfitting)
• Idea 2: use questions in the form of (x1>t?), where t is a threshold.
There are fast ways to try all(?) t.

H ( y∣x i >t ? )=p ( x i >t ) H ( y∣x i > t )+ p( x i ≤t ) H ( y∣x i ≤t )

I ( y∣x i >t ? )=H ( y )−H ( y∣x i >t ? )
What does the feature space
look like?

Axis-parallel cuts
 Conclusions
• Decision trees are popular tools for data
mining
– Easy to understand
– Easy to implement
– Easy to use
– Computationally cheap
• Overfitting might happen
• We used decision trees for classification
(predicting a categorical output from
categorical or real inputs)
 What you should know
• Trees for classification
• Top-down tree construction
algorithm
• Information gain
• Overfitting
• Pruning
• Real-valued features