Note to other teachers and users of these slides.

Andrew would be delighted if you found this source

material useful in giving your own lectures. Feel free to
use these slides verbatim, or to modify them to fit your
own needs. PowerPoint originals are available. If you
make use of a significant portion of these slides in your
own lecture, please include this message, or the
following link to the source repository of Andrew’s
tutorials: http://www.cs.cmu.edu/~awm/tutorials .
Comments and corrections gratefully received.

Decision Trees
Andrew W. Moore
Professor
School of Computer Science
Carnegie Mellon University
www.cs.cmu.edu/~awm
[email protected]
412-268-7599

Copyright © Andrew W. Moore Slide 1

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 2

 Here is a dataset
age employmenteducation edunum
marital … job relation race gender hours_worked
country wealth
…
39 State_gov Bachelors 13 Never_married
… Adm_clericalNot_in_family
White Male 40 United_States
poor
51 Self_emp_not_inc
Bachelors 13 Married … Exec_managerial
Husband White Male 13 United_States
poor
39 Private HS_grad 9 Divorced … Handlers_cleaners
Not_in_family
White Male 40 United_States
poor
54 Private 11th 7 Married … Handlers_cleaners
Husband Black Male 40 United_States
poor
28 Private Bachelors 13 Married … Prof_specialty
Wife Black Female 40 Cuba poor
38 Private Masters 14 Married … Exec_managerial
Wife White Female 40 United_States
poor
50 Private 9th 5 Married_spouse_absent
… Other_service
Not_in_family
Black Female 16 Jamaica poor
52 Self_emp_not_inc
HS_grad 9 Married … Exec_managerial
Husband White Male 45 United_States
rich
31 Private Masters 14 Never_married
… Prof_specialty
Not_in_family
White Female 50 United_States
rich
42 Private Bachelors 13 Married … Exec_managerial
Husband White Male 40 United_States
rich
37 Private Some_college
10 Married … Exec_managerial
Husband Black Male 80 United_States
rich
30 State_gov Bachelors 13 Married … Prof_specialty
Husband Asian Male 40 India rich
24 Private Bachelors 13 Never_married
… Adm_clericalOwn_child White Female 30 United_States
poor
33 Private Assoc_acdm12 Never_married
… Sales Not_in_family
Black Male 50 United_States
poor
41 Private Assoc_voc 11 Married … Craft_repairHusband Asian Male 40 *MissingValue*
rich
34 Private 7th_8th 4 Married … Transport_moving
Husband Amer_Indian
M ale 45 Mexico poor
26 Self_emp_not_inc
HS_grad 9 Never_married
… Farming_fishing
Own_child White Male 35 United_States
poor
33 Private HS_grad 9 Never_married
… Machine_op_inspct
Unmarried White Male 40 United_States
poor
38 Private 11th 7 Married … Sales Husband White Male 50 United_States
poor
44 Self_emp_not_inc
Masters 14 Divorced … Exec_managerial
Unmarried White Female 45 United_States
rich
41 Private Doctorate 16 Married … Prof_specialty
Husband White Male 60 United_States
rich
: : : : : : : : : : : : :

48,000 records, 16 attributes [Kohavi 1995]

Copyright © Andrew W. Moore Slide 3

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 4

 Classification
• A Major Data Mining Operation
• Give one attribute (e.g wealth), try to
predict the value of new people’s wealths by
means of some of the other available
attributes.
• Applies to categorical outputs

• Categorical attribute: an attribute which takes on two or more

discrete values. Also known as a symbolic attribute.
• Real attribute: a column of real numbers

Copyright © Andrew W. Moore Slide 5

 Today’s lecture
• Information Gain for measuring association
between inputs and outputs
• Learning a decision tree classifier from data

Copyright © Andrew W. Moore Slide 6

 About this dataset
• It is a tiny subset of the 1990 US Census.
• It is publicly available online from the UCI
Machine Learning Datasets repository

Copyright © Andrew W. Moore Slide 7

 What can you do with a dataset?
• Well, you can look at histograms…

Gender

Marital
Status

Copyright © Andrew W. Moore Slide 8

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 9

 Contingency Tables
• A better name for a histogram:
A One-dimensional Contingency Table
• Recipe for making a k-dimensional
contingency table:
1. Pick k attributes from your dataset. Call them
a1,a2, … ak.
2. For every possible combination of values,
a1,=x1, a2,=x2,… ak,=xk ,record how frequently
that combination occurs
Fun fact: A database person would call this a “k-dimensional datacube”

Copyright © Andrew W. Moore Slide 10

 A 2-d Contingency Table
• For each pair of
values for
attributes
(agegroup,wealth)
we can see how
many records
match.

Copyright © Andrew W. Moore Slide 11

 A 2-d Contingency Table

• Easier to
appreciate
graphically

Copyright © Andrew W. Moore Slide 12

 A 2-d Contingency Table

• Easier to see
“interesting”
things if we
stretch out the
histogram
bars

Copyright © Andrew W. Moore Slide 13

 A bigger 2-d contingency table

Copyright © Andrew W. Moore Slide 14

 3-d contingency tables
• These are harder to look at!

50s
40s
30s
20s
Po
o r
Ric
h

Male

Female

Copyright © Andrew W. Moore Slide 15

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 16

 On-Line Analytical
Processing (OLAP)
• Software packages and database add-ons to do
this are known as OLAP tools
• They usually include point and click navigation to
view slices and aggregates of contingency tables
• They usually include nice histogram visualization

Copyright © Andrew W. Moore Slide 17

 Time to stop and think
• Why would people want to look at
contingency tables?

Copyright © Andrew W. Moore Slide 18

 Let’s continue to think
• With 16 attributes, how many 1-d
contingency tables are there?
• How many 2-d contingency tables?
• How many 3-d tables?
• With 100 attributes how many 3-d tables
are there?

Copyright © Andrew W. Moore Slide 19

 Let’s continue to think
• With 16 attributes, how many 1-d
contingency tables are there? 16
• How many 2-d contingency tables? 16-
choose-2 = 16 * 15 / 2 = 120
• How many 3-d tables? 560
• With 100 attributes how many 3-d tables
are there? 161,700

Copyright © Andrew W. Moore Slide 20

 Manually looking at contingency
tables
• Looking at one contingency table: can be as much
fun as reading an interesting book
• Looking at ten tables: as much fun as watching CNN
• Looking at 100 tables: as much fun as watching an
infomercial
• Looking at 100,000 tables: as much fun as a three-
week November vacation in Duluth with a dying weasel.

Copyright © Andrew W. Moore Slide 21

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 22

 Data Mining
• Data Mining is all about automating the
process of searching for patterns in the
data.

Which patterns are interesting?

Which might be mere illusions?
And how can they be exploited?

Copyright © Andrew W. Moore Slide 23

 Data Mining
• Data Mining is all about automating the
process of searching for patterns in the
data.
That’s
what we’ll
Which patterns are interesting? look at
right now.
Which might be mere illusions?
And the
And how can they be exploited? answer will
turn out to
be the
engine
that drives
decision
Copyright © Andrew W. Moore Slide 24
 Deciding whether a pattern is
interesting
• We will use information theory
• A very large topic, originally used for
compressing signals
• But more recently used for data mining…

Copyright © Andrew W. Moore Slide 25

 Deciding whether a pattern is
interesting
• We will use information theory
• A very large topic, originally used for
compressing signals
• But more recently used for data mining…

(The topic of Information Gain will now be

discussed, but you will find it in a separate
Andrew Handout)

Copyright © Andrew W. Moore Slide 26

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 27

 Searching for High Info Gains
• Given something (e.g. wealth) you are trying to
predict, it is easy to ask the computer to find
which attribute has highest information gain for it.

Copyright © Andrew W. Moore Slide 28

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 29

 Learning Decision Trees
• A Decision Tree is a tree-structured plan of
a set of attributes to test in order to predict
the output.
• To decide which attribute should be tested
first, simply find the one with the highest
information gain.
• Then recurse…

Copyright © Andrew W. Moore Slide 30

 A small dataset: Miles Per Gallon
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

40 bad
bad
8
6
high
medium
high
medium
high
medium
low
medium
70to74
70to74
america
america
Records bad
bad
4
4
low
low
medium
medium
low
low
medium
low
70to74
70to74
asia
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

From the UCI repository (thanks to Ross Quinlan)

Copyright © Andrew W. Moore Slide 31

 Suppose we want to
predict MPG.

Look at all
the
information
gains…

Copyright © Andrew W. Moore Slide 32

 A Decision Stump

Copyright © Andrew W. Moore Slide 33

 Recursion Step
Records
in which
cylinders
=4

Records
in which
cylinders
=5
Take the And partition it
Original according
Dataset.. Records
to the value of
in which
the attribute cylinders
we split on =6

Records
in which
cylinders
=8

Copyright © Andrew W. Moore Slide 34

 Recursion Step

Build tree from Build tree from Build tree from Build tree from
These records.. These records.. These records.. These records..

Records in
Records in which cylinders
which cylinders =8
=6
Records in
Records in
which cylinders
which cylinders
=5
=4

Copyright © Andrew W. Moore Slide 35

 Second level of tree

Recursively build a tree from the seven (Similar recursion in the

records in which there are four cylinders and other cases)
the maker was based in Asia

Copyright © Andrew W. Moore Slide 36

 The final tree

Copyright © Andrew W. Moore Slide 37

 Base Case
One

Don’t split a
node if all
matching
records have
the same
output value

Copyright © Andrew W. Moore Slide 38

 Base Case
Two

Don’t split a
node if none
of the
attributes can
create
multiple non-
empty
children

Copyright © Andrew W. Moore Slide 39

 Base Case Two:
No attributes
can distinguish

Copyright © Andrew W. Moore Slide 40

 Base Cases
• Base Case One: If all records in current data subset have
the same output then don’t recurse
• Base Case Two: If all records have exactly the same set of
input attributes then don’t recurse

Copyright © Andrew W. Moore Slide 41

 Base Cases: An idea
• Base Case One: If all records in current data subset have
the same output then don’t recurse
• Base Case Two: If all records have exactly the same set of
input attributes then don’t recurse

Proposed Base Case 3:

If all attributes have zero information

gain then don’t recurse

•Is this a good idea?

Copyright © Andrew W. Moore Slide 42

 The problem with Base Case 3
a b y
0 0 0 y = a XOR b
0 1 1
1 0 1
1 1 0

The resulting decision

The information gains: tree:

Copyright © Andrew W. Moore Slide 43

 If we omit Base Case 3:
a b y
0 0 0 y = a XOR b
0 1 1
1 0 1
1 1 0

The resulting decision tree:

Copyright © Andrew W. Moore Slide 44

 Basic Decision Tree Building
Summarized
BuildTree(DataSet,Output)
• If all output values are the same in DataSet, return a leaf node that
says “predict this unique output”
• If all input values are the same, return a leaf node that says “predict
the majority output”
• Else find attribute X with highest Info Gain
• Suppose X has nX distinct values (i.e. X has arity nX).
• Create and return a non-leaf node with nX children.
• The i’th child should be built by calling
BuildTree(DSi,Output)
Where DSi built consists of all those records in DataSet for which X = ith
distinct value of X.

Copyright © Andrew W. Moore Slide 45

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 46

 Training Set Error
• For each record, follow the decision tree to
see what it would predict
For what number of records does the decision
tree’s prediction disagree with the true value in
the database?
• This quantity is called the training set error.
The smaller the better.

Copyright © Andrew W. Moore Slide 47

 MPG Training
error

Copyright © Andrew W. Moore Slide 48

 MPG Training
error

Copyright © Andrew W. Moore Slide 49

 MPG Training
error

Copyright © Andrew W. Moore Slide 50

 Stop and reflect: Why are we
doing this learning anyway?
• It is not usually in order to predict the
training data’s output on data we have
already seen.

Copyright © Andrew W. Moore Slide 51

 Stop and reflect: Why are we
doing this learning anyway?
• It is not usually in order to predict the
training data’s output on data we have
already seen.
• It is more commonly in order to predict the
output value for future data we have not yet
seen.

Copyright © Andrew W. Moore Slide 52

 Stop and reflect: Why are we
doing this learning anyway?
• It is not usually in order to predict the
training data’s output on data we have
already seen.
• It is more commonly in order to predict the
output value for future data we have not yet
seen.

Warning: A common data mining misperception is that the

above two bullets are the only possible reasons for learning.
There are at least a dozen others.

Copyright © Andrew W. Moore Slide 53

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 54

 Test Set Error
• Suppose we are forward thinking.
• We hide some data away when we learn the
decision tree.
• But once learned, we see how well the tree
predicts that data.
• This is a good simulation of what happens
when we try to predict future data.
• And it is called Test Set Error.

Copyright © Andrew W. Moore Slide 55

 MPG Test set
error

Copyright © Andrew W. Moore Slide 56

 MPG Test set
error

The test set error is much worse than the

training set error…
…why?

Copyright © Andrew W. Moore Slide 57

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 58

 An artificial example
• We’ll create a training dataset
Output y = copy of e,
Five inputs, all bits, are Except a random 25%
generated in all 32 possible of the records have y
combinations set to the opposite of e

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

0 0 0 1 0 0
0 0 0 1 1 1
0 0 1 0 0 1
: : : : : :
1 1 1 1 1 1

Copyright © Andrew W. Moore Slide 59

 In our artificial example
• Suppose someone generates a test set
according to the same method.
• The test set is identical, except that some of
the y’s will be different.
• Some y’s that were corrupted in the training
set will be uncorrupted in the testing set.
• Some y’s that were uncorrupted in the
training set will be corrupted in the test set.

Copyright © Andrew W. Moore Slide 60

 Building a tree with the artificial
training set
• Suppose we build a full tree (we always split until base case 2)
Root

e=0 e=1

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

25% of these leaf node labels will be corrupted

Copyright © Andrew W. Moore Slide 61

 Training set error for our artificial
tree
All the leaf nodes contain exactly one record and so…

• We would have a training set error

of zero

Copyright © Andrew W. Moore Slide 62

 Testing the tree with the test set
1/4 of the tree nodes 3/4 are fine
are corrupted
1/4 of the test set 1/16 of the test set will 3/16 of the test set will
records are be correctly predicted be wrongly predicted
corrupted for the wrong reasons because the test record is
corrupted
3/4 are fine 3/16 of the test 9/16 of the test
predictions will be predictions will be fine
wrong because the
tree node is corrupted

In total, we expect to be wrong on 3/8 of the test set predictions

Copyright © Andrew W. Moore Slide 63

 What’s this example shown us?
• This explains the discrepancy between
training and test set error
• But more importantly… …it indicates there’s
something we should do about it if we want
to predict well on future data.

Copyright © Andrew W. Moore Slide 64

 Suppose we had less data
• Let’s not look at the irrelevant bits
Output y = copy of e, except a
random 25% of the records
These bits are hidden 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
32 records

0 0 0 1 0 0
0 0 0 1 1 1
0 0 1 0 0 1
: : : : : :
1 1 1 1 1 1

What decision tree would we learn now?

Copyright © Andrew W. Moore Slide 65
 Without access to the irrelevant bits…
Root

e=0 e=1
These nodes will be unexpandable

Copyright © Andrew W. Moore Slide 66

 Without access to the irrelevant bits…
Root

e=0 e=1
These nodes will be unexpandable

In about 12 of In about 12 of
the 16 records the 16 records
in this node the in this node the
output will be 0 output will be 1

So this will So this will

almost certainly almost certainly
predict 0 predict 1
Copyright © Andrew W. Moore Slide 67
 Without access to the irrelevant bits…
almost certainly almost certainly all
Root none of the tree are fine
nodes are
corrupted
e=0 e=1
1/4 of the test n/a 1/4 of the test set
set records will be wrongly
are corrupted predicted because
the test record is
corrupted
3/4 are fine n/a 3/4 of the test
predictions will be
fine

In total, we expect to be wrong on only 1/4 of the test set predictions

Copyright © Andrew W. Moore Slide 68

 Overfitting
• Definition: If your machine learning
algorithm fits noise (i.e. pays attention to
parts of the data that are irrelevant) it is
overfitting.
• Fact (theoretical and empirical): If your
machine learning algorithm is overfitting
then it may perform less well on test set
data.

Copyright © Andrew W. Moore Slide 69

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 70

 Avoiding overfitting
• Usually we do not know in advance which
are the irrelevant variables
• …and it may depend on the context
For example, if y = a AND b then b is an irrelevant
variable only in the portion of the tree in which a=0

But we can use simple statistics to

warn us that we might be
overfitting.

Copyright © Andrew W. Moore Slide 71

 Consider this
split

Copyright © Andrew W. Moore Slide 72

 A chi-squared test

• Suppose that mpg was completely uncorrelated with

maker.
• What is the chance we’d have seen data of at least this
apparent level of association anyway?

Copyright © Andrew W. Moore Slide 73

 A chi-squared test

• Suppose that mpg was completely uncorrelated with

maker.
• What is the chance we’d have seen data of at least this
apparent level of association anyway?
By using a particular kind of chi-squared test, the
answer is 13.5%.

Copyright © Andrew W. Moore Slide 74

 Using Chi-squared to avoid
overfitting
• Build the full decision tree as before.
• But when you can grow it no more, start to
prune:
• Beginning at the bottom of the tree, delete
splits in which pchance > MaxPchance.
• Continue working you way up until there are no
more prunable nodes.

MaxPchance is a magic parameter you must specify to the decision tree,

indicating your willingness to risk fitting noise.

Copyright © Andrew W. Moore Slide 75

 Pruning example
• With MaxPchance = 0.1, you will see the
following MPG decision tree:

Note the improved

test set accuracy
compared with the
unpruned tree

Copyright © Andrew W. Moore Slide 76

 MaxPchance
• Good news: The decision tree can automatically adjust
its pruning decisions according to the amount of apparent
noise and data.
• Bad news: The user must come up with a good value
of MaxPchance. (Note, Andrew usually uses 0.05, which is
his favorite value for any magic parameter).
• Good news: But with extra work, the best MaxPchance
value can be estimated automatically by a technique called
cross-validation.

Copyright © Andrew W. Moore Slide 77

 MaxPchance
• Technical note (dealt with in other lectures):
MaxPchance is a regularization parameter.
Expected Test set
Error

Decreasing Increasing
MaxPchance

High Bias High Variance

Copyright © Andrew W. Moore Slide 78

 The simplest tree
• Note that this pruning is heuristically trying
to find
The simplest tree structure for which all within-leaf-
node disagreements can be explained by chance
• This is not the same as saying “the simplest
classification scheme for which…”
• Decision trees are biased to prefer classifiers
that can be expressed as trees.

Copyright © Andrew W. Moore Slide 79

 Expressiveness of Decision Trees
• Assume all inputs are Boolean and all outputs are
Boolean.
• What is the class of Boolean functions that are
possible to represent by decision trees?
• Answer: All Boolean functions.
Simple proof:
1. Take any Boolean function
2. Convert it into a truth table
3. Construct a decision tree in which each row of the truth table
corresponds to one path through the decision tree.

Copyright © Andrew W. Moore Slide 80

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 81

 Real-Valued inputs
• What should we do if some of the inputs are
real-valued?
mpg cylinders displacementhorsepower weight acceleration modelyear maker

good 4 97 75 2265 18.2 77 asia

bad 6 199 90 2648 15 70 america
bad 4 121 110 2600 12.8 77 europe
bad 8 350 175 4100 13 73 america
bad 6 198 95 3102 16.5 74 america
bad 4 108 94 2379 16.5 73 asia
bad 4 113 95 2228 14 71 asia
bad 8 302 139 3570 12.8 78 america
: : : : : : : :
: : : : : : : :
: : : : : : : :
good 4 120 79 2625 18.6 82 america
bad 8 455 225 4425 10 70 america
good 4 107 86 2464 15.5 76 europe
bad 5 131 103 2830 15.9 78 europe

Idea One: Branch on each possible real value

Copyright © Andrew W. Moore Slide 82

 “One branch for each numeric
value” idea:

Hopeless: with such high branching factor will shatter

the dataset and over fit

Note pchance is 0.222 in the above…if MaxPchance

was 0.05 that would end up pruning away to a single
root node.
Copyright © Andrew W. Moore Slide 83
 A better idea: thresholded splits
• Suppose X is real valued.
• Define IG(Y|X:t) as H(Y) - H(Y|X:t)
• Define H(Y|X:t) =
H(Y|X < t) P(X < t) + H(Y|X >= t) P(X >= t)
• IG(Y|X:t) is the information gain for predicting Y if all
you know is whether X is greater than or less than t
• Then define IG*(Y|X) = maxt IG(Y|X:t)
• For each real-valued attribute, use IG*(Y|X)
for assessing its suitability as a split
Copyright © Andrew W. Moore Slide 84
 Computational Issues
• You can compute IG*(Y|X) in time
R log R + 2 R ny
• Where
R is the number of records in the node under consideration
ny is the arity (number of distinct values of) Y

How?
Sort records according to increasing values of X. Then create a 2x ny
contingency table corresponding to computation of IG(Y|X:x min). Then
iterate through the records, testing for each threshold between adjacent
values of X, incrementally updating the contingency table as you go. For a
minor additional speedup, only test between values of Y that differ.

Copyright © Andrew W. Moore Slide 85

 Example with
MPG

Copyright © Andrew W. Moore Slide 86

 Unpruned
tree using
reals

Copyright © Andrew W. Moore Slide 87

 Pruned tree using reals

Copyright © Andrew W. Moore Slide 88

 LearnUnprunedTree(X,Y)
Input: X a matrix of R rows and M columns where X ij = the value of the j’th attribute in the i’th input datapoint. Each
column consists of either all real values or all categorical values.
Input: Y a vector of R elements, where Y i = the output class of the i’th datapoint. The Y i values are categorical.
Output: An Unpruned decision tree

If all records in X have identical values in all their attributes (this includes the case where R<2), return a Leaf Node
predicting the majority output, breaking ties randomly. This case also includes
If all values in Y are the same, return a Leaf Node predicting this value as the output
Else
For j = 1 .. M
If j’th attribute is categorical
IGj = IG(Y|Xj)
Else (j’th attribute is real-valued)
IGj = IG*(Y|Xj) from about four slides back
Let j* = argmaxj IGj (this is the splitting attribute we’ll use)
If j* is categorical then
For each value v of the j’th attribute
Let Xv = subset of rows of X in which Xij = v. Let Yv = corresponding subset of Y
Let Childv = LearnUnprunedTree(Xv,Yv)
Return a decision tree node, splitting on j’th attribute. The number of children equals the number of values of
the j’th attribute, and the v’th child is Child v
Else j* is real-valued and let t be the best split threshold
Let XLO = subset of rows of X in which Xij <= t. Let YLO = corresponding subset of Y
Let ChildLO = LearnUnprunedTree(XLO,YLO)
Let XHI = subset of rows of X in which Xij > t. Let YHI = corresponding subset of Y
Let ChildHI = LearnUnprunedTree(XHI,YHI)
Return a decision tree node, splitting on j’th attribute. It has two children corresponding to whether the j’th
attribute is above or below the given threshold.

Copyright © Andrew W. Moore Slide 89

 LearnUnprunedTree(X,Y)
Input: X a matrix of R rows and M columns where X ij = the value of the j’th attribute in the i’th input datapoint. Each
column consists of either all real values or all categorical values.
Input: Y a vector of R elements,Twhere
hingsYi to nooutput
= the te: class of the i’th datapoint. The Y i values are categorical.
Output: An Unpruned decision tree
B elow the root nod
e, there is no poin
testing categorica t
If all records in X have identical values in all their attributes l att(this
ribuincludes
te s that have
the case where R<2), return a Leaf Node
predicting the majority output, alrebreaking
ady beties en randomly.
split upon further
This case also includes
If all values in Y are the same,T his isa Leaf upoutput
the tree.
return becausepredicting
Node
all thethis value as the
values of that
Else attribute will be th
e same and IG mu
For j = 1 .. M therefore be zero. st
If j’th attribute is categorical
Buj)t it’s worth rete
IGj = IG(Y|X
sting real-valued
attr
Else (j’th attribute is ib utes, since they m
real-valued)
IGj = IG*(Y|X
valuj)efrom four slides backay have different
s beabout
low the b ary split
Let j* = argmaxj IGj (this is the splitting attributein use) , and may
bene fit from splittingwe’ll further.
If j* is categorical then
For each value T voof a chj’th
the ievattribute
e the above optim
Let X =sh
v
ould passofdXow
subset of rows in which Xij = v. Let izat
Yv io
=n , you
corresponding subset of Y
n th ro u g h th e recursion a
Let Childcu rrent active set of
v = LearnUnprunedTree(Xv,Yv)

Return a decision tree node, splitting on j’th attribute.at tributeThe s. number of children equals the number of values of
Pedanand
the j’th attribute, ticthe
dev’th child is Child
tail:
v
a third termination
Else j* is real-valuedco andn let t be the best split threshold
dition should occu
Let XLO = subset of rows of X in which Xij <=r t.ifLet
thYeLOb=es t split
corresponding subset of Y
at tr ib u te p u ts allLO,YitLOs) records in e
Let ChildLO = LearnUnprunedTree(X
ch ild (n xactly one
Let XHI = subset of rowsoof teX th
in at this
which Xij > mt.eLet
ansYHIit=acorresponding
nd all other subset of Y
att rib u te s
Let ChildHI = LearnUnprunedTree(X have IG HI
,Y= HI
)0).
Return a decision tree node, splitting on j’th attribute. It has two children corresponding to whether the j’th
attribute is above or below the given threshold.

Copyright © Andrew W. Moore Slide 90

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real Valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 91

 Warning: unlike
what went befo
trade”: not part re, this is an ed
of the official D itorial “trick of th
ecision Tree alg e
orithm.

Binary categorical splits

• One of Andrew’s
favorite tricks Example:

• Allow splits of the

following form
Root

Attribute Attribute
equals doesn’t
value equal value

Copyright © Andrew W. Moore Slide 92

 Machine Learning Datasets
What is Classification?

Outline Contingency Tables

OLAP (Online Analytical Processing)
What is Data Mining?
Searching for High Information Gain
Learning an unpruned decision tree recursively
Training Set Error
Test Set Error
Overfitting
Avoiding Overfitting
Information Gain of a real valued input
Building Decision Trees with real valued Inputs
Andrew’s homebrewed hack: Binary Categorical Splits
Example Decision Trees

Copyright © Andrew W. Moore Slide 93

 Predicting age
from census

Copyright © Andrew W. Moore Slide 94

 Predicting
wealth from
census

Copyright © Andrew W. Moore Slide 95

 Predicting gender from census

Copyright © Andrew W. Moore Slide 96

 Conclusions
• Decision trees are the single most popular
data mining tool
• Easy to understand
• Easy to implement
• Easy to use
• Computationally cheap
• It’s possible to get in trouble with overfitting
• They do classification: predict a categorical
output from categorical and/or real inputs
Copyright © Andrew W. Moore Slide 97
 What you should know
• What’s a contingency table?
• What’s information gain, and why we use it
• The recursive algorithm for building an
unpruned decision tree
• What are training and test set errors
• Why test set errors can be bigger than
training set
• Why pruning can reduce test set error
• How to exploit real-valued inputs
Copyright © Andrew W. Moore Slide 98
 What we haven’t discussed
• It’s easy to have real-valued outputs too---these are called
Regression Trees*
• Bayesian Decision Trees can take a different approach to
preventing overfitting
• Computational complexity (straightforward and cheap) *
• Alternatives to Information Gain for splitting nodes
• How to choose MaxPchance automatically *
• The details of Chi-Squared testing *
• Boosting---a simple way to improve accuracy *

* = discussed in other Andrew

lectures

Copyright © Andrew W. Moore Slide 99

 For more information
• Two nice books
• L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone.
Classification and Regression Trees. Wadsworth, Belmont,
CA, 1984.
• C4.5 : Programs for Machine Learning (Morgan Kaufmann
Series in Machine Learning) by J. Ross Quinlan
• Dozens of nice papers, including
• Learning Classification Trees, Wray Buntine, Statistics and
Computation (1992), Vol 2, pages 63-73
• Kearns and Mansour, On the Boosting Ability of Top-Down
Decision Tree Learning Algorithms, STOC: ACM Symposium
on Theory of Computing, 1996“
• Dozens of software implementations available on the web for free and
commercially for prices ranging between $50 - $300,000
Copyright © Andrew W. Moore Slide 100
 Discussion
• Instead of using information gain, why not choose the
splitting attribute to be the one with the highest prediction
accuracy?
• Instead of greedily, heuristically, building the tree, why not
do a combinatorial search for the optimal tree?
• If you build a decision tree to predict wealth, and marital
status, age and gender are chosen as attributes near the
top of the tree, is it reasonable to conclude that those three
inputs are the major causes of wealth?
• ..would it be reasonable to assume that attributes not
mentioned in the tree are not causes of wealth?
• ..would it be reasonable to assume that attributes not
mentioned in the tree are not correlated with wealth?
Copyright © Andrew W. Moore Slide 101