Upload
sophia-jenkins
View
221
Download
1
Tags:
Embed Size (px)
Citation preview
1
Decision Tree Learning
Soongsil University, SeoulGun Ho Lee
2
Decision Tree Learning
• Introduction• Decision Tree Representation• Appropriate Problems for Decision Tree Learni
ng• Basic Algorithm• Hypothesis Space Search in Decision Tree Lea
rning• Inductive Bias in Decision Tree Learning• Issues in Decision Tree Learning• Summary
3
Tree leaning Task
Apply
Model
Induction
Deduction
Learn
Model (tree)
Model(Tree)
Tid Attrib1 Attrib2 Attrib3 Class
1 Yes Large 125K No
2 No Medium 100K No
3 No Small 70K No
4 Yes Medium 120K No
5 No Large 95K Yes
6 No Medium 60K No
7 Yes Large 220K No
8 No Small 85K Yes
9 No Medium 75K No
10 No Small 90K Yes 10
Tid Attrib1 Attrib2 Attrib3 Class
11 No Small 55K ?
12 Yes Medium 80K ?
13 Yes Large 110K ?
14 No Small 95K ?
15 No Large 67K ? 10
Test Set
Learningalgorithm
Training Set
4
Example of a Decision Tree
Tid Refund MaritalStatus
TaxableIncome Cheat
1 Yes Single 125K No
2 No Married 100K No
3 No Single 70K No
4 Yes Married 120K No
5 No Divorced 95K Yes
6 No Married 60K No
7 Yes Divorced 220K No
8 No Single 85K Yes
9 No Married 75K No
10 No Single 90K Yes10
categoric
al
categoric
al
continuous
class
Refund
MarSt
TaxInc
YESNO
NO
NO
Yes No
Married Single, Divorced
< 80K > 80K
Splitting Attributes
Training Data Model: Decision Tree
5
Another Example of Decision Tree
Tid Refund MaritalStatus
TaxableIncome Cheat
1 Yes Single 125K No
2 No Married 100K No
3 No Single 70K No
4 Yes Married 120K No
5 No Divorced 95K Yes
6 No Married 60K No
7 Yes Divorced 220K No
8 No Single 85K Yes
9 No Married 75K No
10 No Single 90K Yes10
categoric
al
categoric
al
continuous
classMarSt
Refund
TaxInc
YESNO
NO
NO
Yes No
Married Single,
Divorced
< 80K > 80K
There could be more than one tree that fits the same data!
6
Decision Tree Classification Task
Apply
Model
Induction
Deduction
Learn
Model
Model
Tid Attrib1 Attrib2 Attrib3 Class
1 Yes Large 125K No
2 No Medium 100K No
3 No Small 70K No
4 Yes Medium 120K No
5 No Large 95K Yes
6 No Medium 60K No
7 Yes Large 220K No
8 No Small 85K Yes
9 No Medium 75K No
10 No Small 90K Yes 10
Tid Attrib1 Attrib2 Attrib3 Class
11 No Small 55K ?
12 Yes Medium 80K ?
13 Yes Large 110K ?
14 No Small 95K ?
15 No Large 67K ? 10
Test Set
TreeInductionalgorithm
Training Set
Decision Tree
7
Apply Model to Test Data
Refund
MarSt
TaxInc
YESNO
NO
NO
Yes No
Married Single, Divorced
< 80K > 80K
Refund Marital Status
Taxable Income Cheat
No Married 80K ? 10
Test DataStart from the root of tree.
8
Apply Model to Test Data
Refund
MarSt
TaxInc
YESNO
NO
NO
Yes No
Married Single, Divorced
< 80K > 80K
Refund Marital Status
Taxable Income Cheat
No Married 80K ? 10
Test Data
9
Apply Model to Test Data
Refund
MarSt
TaxInc
YESNO
NO
NO
Yes No
Married Single, Divorced
< 80K > 80K
Refund Marital Status
Taxable Income Cheat
No Married 80K ? 10
Test Data
10
Apply Model to Test Data
Refund
MarSt
TaxInc
YESNO
NO
NO
Yes No
Married Single, Divorced
< 80K > 80K
Refund Marital Status
Taxable Income Cheat
No Married 80K ? 10
Test Data
11
Apply Model to Test Data
Refund
MarSt
TaxInc
YESNO
NO
NO
Yes No
Married Single, Divorced
< 80K > 80K
Refund Marital Status
Taxable Income Cheat
No Married 80K ? 10
Test Data
12
Apply Model to Test Data
Refund
MarSt
TaxInc
YESNO
NO
NO
Yes No
Married Single, Divorced
< 80K > 80K
Refund Marital Status
Taxable Income Cheat
No Married 80K ? 10
Test Data
Assign Cheat to “No”
13
Decision Tree Classification Task
Apply
Model
Induction
Deduction
Learn
Model
Model
Tid Attrib1 Attrib2 Attrib3 Class
1 Yes Large 125K No
2 No Medium 100K No
3 No Small 70K No
4 Yes Medium 120K No
5 No Large 95K Yes
6 No Medium 60K No
7 Yes Large 220K No
8 No Small 85K Yes
9 No Medium 75K No
10 No Small 90K Yes 10
Tid Attrib1 Attrib2 Attrib3 Class
11 No Small 55K ?
12 Yes Medium 80K ?
13 Yes Large 110K ?
14 No Small 95K ?
15 No Large 67K ? 10
Test Set
TreeInductionalgorithm
Training Set
Decision Tree
14
Overview• One of the most widely used and practical methods for
inductive inference over supervised data
• It approximates discrete-valued functions (as opposed to continuous)
• It is robust to noisy data
• Decision trees can represent any discrete function on discrete features
• It is also efficient for processing large amounts of data, so is often used in data mining applications
• Decision Tree Learners’ bias typically prefers small tree over larger ones
15
Decision Trees
• Tree-based classifiers for instances represented as feature-vectors. Nodes test features, there is one branch for each value of the feature, and leaves specify the category.
• Can represent arbitrary conjunction and disjunction. Can represent any classification function over discrete feature vectors.
• Can be rewritten as a set of rules, i.e. disjunctive normal form (DNF).– red circle → pos– red circle → A blue → B; red square → B green → C; red triangle → C
color
red blue green
shape
circlesquare triangle
neg pos
pos neg neg
color
red bluegreen
shape
circle square triangle
B C
A B C
16
Properties of Decision Tree Learning
• Continuous (real-valued) features can be handled by allowing nodes to split a real valued feature into two ranges based on a threshold (e.g. length < 3 and length 3)
• Classification trees have discrete class labels at the leaves, regression trees allow real-valued outputs at the leaves.
• Algorithms for finding consistent trees are efficient for processing large amounts of training data for data mining tasks.
• Methods developed for handling noisy training data (both class and feature noise).
• Methods developed for handling missing feature values.
17
What makes a good tree?
• Not too small – need to include enough attributes to handle possibly subtle distinctions in data
• Not too big
- computational efficiency (avoid redundant, spurious attributes)
- avoid over-fitting training examples (noisy, scarce data)
Occam’s Razor: find simplest hypothesis (tree) that is consistent with all observations
inductive bias – small trees, with informative nodes near root
18
Basic Algorithm
• 가능한 모든 decision trees space 에서의 top-down, greedy search
• Training examples 를 가장 잘 분류할 수 있는 attribute 를 루트에 둔다 .
• Entropy, Information gain
19
Top-Down Decision Tree Induction
• Recursively build a tree top-down by divide and conquer.
Example: <big, red, circle>: + <small, red, circle>: + <small, red, square>: <big, blue, circle>:
<big, red, circle>: + <small, red, circle>: +<small, red, square>:
color
red bluegreen
20
shape
circlesquare
triangle
Top-Down Decision Tree Induction
• Recursively build a tree top-down by divide and conquer.
<big, red, circle>: + <small, red, circle>: +<small, red, square>:
color
red bluegreen
<big, red, circle>: + <small, red, circle>: +
pos
<small, red, square>:
negpos
<big, blue, circle>: neg neg
Example: <big, red, circle>: + <small, red, circle>: + <small, red, square>: <big, blue, circle>:
21
Decision Tree Induction Pseudocode
DTree(examples, features) returns a tree If all examples are in one category, return a leaf node with that category label. Else if the set of features is empty, return a leaf node with the category label that is the most common in examples. Else pick a feature F and create a node R for it For each possible value vi of F: Let examplesi be the subset of examples that have value vi for F
Add an out-going edge E to node R labeled with the value vi.
If examplesi is empty then attach a leaf node to edge E labeled with the category that is the most common in examples. else call DTree(examplesi , features – {F}) and attach the resulting tree as the subtree under edge E. Return the subtree rooted at R.
22
The Basic Decision Tree Learning Algorithm
Top-Down Induction of Decision Trees
. This approach is exemplified by the ID3 algorithm and its successor C4.5
23
Tree Induction
• Greedy strategy.– Split the records based on an attribute test that
optimizes certain criterion.
• Issues– Determine how to split the records
• How to specify the attribute test condition?• How to determine the best split?
– Determine when to stop splitting
24
How to determine the Best Split
Income Age
>=10k<10k young old
Customers
fair customersGood customers
25
How to determine the Best Split
• Greedy approach: – Nodes with homogeneous class distribution are
preferred
• Need a measure of node impurity:
High degreeof impurity
Low degreeof impurity
pure
50% red50% green
75% red25% green
100% red0% green
26
Split Selection Method
• Numerical or ordered attributes: Find a split point that separates the (two) classes
(Yes: No: )
30 35
Age
Age < 33
27
Split Selection Method (Contd.)
• Categorical attributes: How to group?
Sport: Truck: Minivan:
(Sport, Truck) -- (Minivan)
(Sport) --- (Truck, Minivan)
(Sport, Minivan) --- (Truck)
Car
Sport, Truck
Minivan
28
Decision Tree Induction
• Many Algorithms:– Hunt’s Algorithm (1960’s, one of the earliest)– ID3(Quinlan 1979), C4.5(Quinlan 1993)– CART– SLIQ (EDBT’96 — Mehta et al.)
• builds an index for each attribute and only class list and the current attribute list reside in memory
– SPRINT (VLDB’96 — J. Shafer et al.)• constructs an attribute list data structure
– RainForest (VLDB’98 — Gehrke, Ramakrishnan & Ganti)• separates the scalability aspects from the criteria that
determine the quality of the tree• builds an AVC-list (attribute, value, class label)
– BOAT • Uses bootstrapping to create several small
samples
29
History of Decision-Tree Research
• Hunt and colleagues use exhaustive search decision-tree methods (CLS) to model human concept learning in the 1960’s.
• In the late 70’s, Quinlan developed ID3 with the information gain heuristic to learn expert systems from examples.
• Simulataneously, Breiman and Friedman and colleagues develop CART (Classification and Regression Trees), similar to ID3.
• In the 1980’s a variety of improvements are introduced to handle noise, continuous features, missing features, and improved splitting criteria. Various expert-system development tools results.
• Quinlan’s updated decision-tree package (C4.5) released in 1993.
• Weka includes Java version of C4.5 called J48.
30
General Structure of Hunt’s Algorithm
• Let Dt be the set of training records that reach a node t
• General Procedure:– If Dt contains records that
belong the same class yt, then t is a leaf node labeled as yt
– If Dt is an empty set, then t is a leaf node labeled by the default class, yd
– If Dt contains records that belong to more than one class, use an attribute test to split the data into smaller subsets. Recursively apply the procedure to each subset.
Tid Refund Marital Status
Taxable Income Cheat
1 Yes Single 125K No
2 No Married 100K No
3 No Single 70K No
4 Yes Married 120K No
5 No Divorced 95K Yes
6 No Married 60K No
7 Yes Divorced 220K No
8 No Single 85K Yes
9 No Married 75K No
10 No Single 90K Yes 10
Dt
?
31
Hunt’s Algorithm
Cheat=No
Refund
Cheat=No Cheat=No
Yes No
Refund
Cheat=No
Yes No
MaritalStatus
Cheat=No
Cheat
Single,Divorced
Married
TaxableIncome
Cheat=No
< 80K >= 80K
Refund
Cheat=No
Yes No
MaritalStatus
Cheat=NoCheat
Single,Divorced
Married
32
What is ID3(Interactive Dichotomizer 3)
• A mathematical algorithm for building the decision tree.
• Invented by J. Ross Quinlan in 1979.• Uses Information Theory invented by Shannon
in 1948.• Information Gain is used to select the most
useful attribute for classification.• Builds the tree from the top down, with no
backtracking.
33
ID3
• ID3 is designed for that case:– many attributes – training set contains many objects– a reasonably good decision tree is required
without much computation– It is generally found to construct simple decision
tree. But it cannot guarantee that the trees is always best.
34
ID3 Algorithm Overview
• Step 1: Choose a random subset of the training set• Step 2: Form a decision tree that correctly classifies all the objects
in the window
• Step 3:
IF the tree gives the correct answer for all the objects in the training
set,
Then the process terminates;
Else a selection of the incorrectly
classified objects is added to the window and go to Step 2
35
ID3 Algorithm
36
Picking a Good Split Feature
• Goal is to have the resulting tree be as small as possible, per Occam’s razor.
• Finding a minimal decision tree (nodes, leaves, or depth) is an NP-hard optimization problem.
• Top-down divide-and-conquer method does a greedy search for a simple tree but does not guarantee to find the smallest.– General lesson in ML: “Greed is good.”
• Want to pick a feature that creates subsets of examples that are relatively “pure” in a single class so they are “closer” to being leaf nodes.
• There are a variety of heuristics for picking a good test, a popular one is based on information gain that originated with the ID3 system of Quinlan (1979).
37
Entropy
• Minimum number of bits of information needed to encode the classification of an arbitrary member of S
– entropy = 0, if all members in the same class– entropy = 1, if |positive examples|=|negative examples|
)(log)(log)( 020121 ppppSEntropy
38
Example – Information Needed
• n = 5• p = 9• The information needed to generate a decision
tree from this window is:
• E(p,n) = = 0.940 bits14
5log
14
5
14
9log
14
922
39
Entropy
• Entropy (disorder, impurity) of a set of examples, S, relative to a binary classification is:
where p1 is the fraction of positive examples in S and p0 is the fraction of negatives.
• If all examples are in one category, entropy is zero (we define 0log(0)=0)
• If examples are equally mixed (p1=p0=0.5), entropy is a maximum of 1.
• For multi-class problems with c categories, entropy generalizes to:
)(log)(log)( 020121 ppppSEntropy
c
iii ppSEntropy
12 )(log)(
40
Entropy Plot for Binary Classification
41
Information Gain
Parent Node, p is split into k partitions;
ni is number of records in partition I
– Measures Reduction in Entropy achieved because of the split.
– Choose the split that achieves maximum GAIN value !!
– Used in ID3 and C4.5
– Disadvantage: Tends to prefer splits that result in large number of partitions, each being small but pure.
k
i
isplit iE
n
npEntropyGAIN
1
)()(
42
Information Gain
• Example:– <big, red, circle>: + <small, red, circle>: +– <small, red, square>: <big, blue, circle>:
2+, 2 : E=1 size
big small
1+,1 1+,1
E=1 E=1
Gain=1(0.51 + 0.51) = 0
2+, 2 : E=1 color
red blue
2+,1 0+,1
E=0.918 E=0
Gain=1(0.750.918 +0.250) = 0.311
2+, 2 : E=1 shape
circle square
2+,1 0+,1
E=0.918 E=0
Gain =1(0.750.918+0.250) = 0.311
43
Training Examples for PlayTennis
Day Outlook 온도 Humidity Wind PlayTennis
D1 Sunny Hot High Weak No
D2 Sunny Hot High Strong No
D3 Overcast Hot High Weak Yes
D4 Rain Mild High Weak Yes
D5 Rain Cool Normal Weak Yes
D6 Rain Cool Normal Strong No
D7 Overcast Cool Normal Strong Yes
D8 Sunny Mild High Weak No
D9 Sunny Cool Normal Weak Yes
D10 Rain Mild Normal Weak Yes
D11 Sunny Mild Normal Strong Yes
D12 Overcast Mild High Strong Yes
D13 Overcast Hot Normal Weak Yes
D14 Rain Mild High Strong No
44
ID3
playdon’t play
pno = 5/14
pyes = 9/14
Impurity = - pyes log2 pyes - pno log2 pno
= - 9/14 log2 9/14 - 5/14 log2 5/14
= 0.94 bits
outlook temperature humidity windy playsunny hot high FALSE nosunny hot high TRUE noovercast hot high FALSE yesrainy mild high FALSE yesrainy cool normal FALSE yesrainy cool normal TRUE noovercast cool normal TRUE yessunny mild high FALSE nosunny cool normal FALSE yesrainy mild normal FALSE yessunny mild normal TRUE yesovercast mild high TRUE yesovercast hot normal FALSE yesrainy mild high TRUE no
45
Selecting the Next Attribute : which attribute is the best classifier?
ID3
46
Example – Branch on outlook
E(outlook) = 0.694
Gain(outlook)=0.940-E(outlook)=0.246bits
Outlook
sunny overcast rain
PTrueNormalMild11
PFalseNormalCool9
NFalseHighMild8
NTrueHighHot2
NFalseHighHot1
PFalseNormalHot13
PTrueHighMild12
PtrueNormalCool7
PFalseHighHot3
NTrueHighMild14
PFalseNormalMild10
NTrueNormalCool6
PFalseNormalCool5
PFalseHighMild4
),(1 45),(1 4
4),(1 45
332211 npEnpEnpE
p3=3,
n3=2,
E(p3, n3)=0.971
p2=4,
n2=0,
E(p2, n2)=0
p1=2,
n1=3,
E(p1, n1)=0.971
47
ID3 play
don’t play
amount of information required to specify class of an example given that it reaches node
0.94 bits
0.0 bits* 4/14
0.97 bits* 5/14
0.97 bits* 5/14
0.98 bits* 7/14
0.59 bits* 7/14
0.92 bits* 6/14
0.81 bits* 4/14
0.81 bits* 8/14
1.0 bits* 4/14
1.0 bits* 6/14
outlook
sunny overcast rainy
+= 0.69 bits
gain: 0.25 bits
+= 0.79 bits
+= 0.91 bits
+= 0.89 bits
gain: 0.15 bits gain: 0.03 bits gain: 0.05 bits
play don't playsunny 2 3
overcast 4 0rainy 3 2
humidity temperature windy
high normal hot mild cool false true
play don't playhot 2 2mild 4 2cool 3 1
play don't playhigh 3 4
normal 6 1
play don't playFALSE 6 2TRUE 3 3
maximal
information
gain
48
ID3 play
don’t playoutlook
sunny overcast rainy
maximal
information
gain
0.97 bits
0.0 bits* 3/5
humidity temperature windy
high normal hot mild cool false true
+= 0.0 bits
gain: 0.97 bits
+= 0.40 bits
gain: 0.57 bits
+= 0.95 bits
gain: 0.02 bits
0.0 bits* 2/5
0.0 bits* 2/5
1.0 bits* 2/5
0.0 bits* 1/5
0.92 bits* 3/5
1.0 bits* 2/5
outlook temperature humidity windy playsunny hot high FALSE nosunny hot high TRUE nosunny mild high FALSE nosunny cool normal FALSE yessunny mild normal TRUE yes
49
ID3 play
don’t playoutlook
sunny overcast rainy
humidity
high normal
outlook temperature humidity windy playrainy mild high FALSE yesrainy cool normal FALSE yesrainy cool normal TRUE norainy mild normal FALSE yesrainy mild high TRUE no
1.0 bits*2/5
temperature windy
hot mild cool false true
+= 0.95 bits
gain: 0.02 bits
+= 0.95 bits
gain: 0.02 bits
+= 0.0 bits
gain: 0.97 bits
humidity
high normal
0.92 bits* 3/5
0.92 bits* 3/5
1.0 bits* 2/5
0.0 bits* 3/5
0.0 bits* 2/5
0.97 bits
50
ID3
play
don’t playoutlook
sunny overcast rainy
windy
false true
humidityhigh
normal
outlook temperature humidity windy playsunny hot high FALSE nosunny hot high TRUE noovercast hot high FALSE yesrainy mild high FALSE yesrainy cool normal FALSE yesrainy cool normal TRUE noovercast cool normal TRUE yessunny mild high FALSE nosunny cool normal FALSE yesrainy mild normal FALSE yessunny mild normal TRUE yesovercast mild high TRUE yesovercast hot normal FALSE yesrainy mild high TRUE no
Yes
NoNo Yes Yes
51
Hypothesis Space Search in Decision Tree Learning
• ID3 can be characterized as searching a space of hypothesis for one that fits the training examples
• The hypothesis space searched by ID3 is the set of possible decisions
• ID3 performs a simple-to-complex, hill-climbing search through this hypothesis space, locally-optimal solution.
52
Hypothesis Space Search in Decision Tree Learning
• Hypothesis space of all decision tree is a complete space of finite discrete-valued functions, relative to the available attributes
• Outputs a single hypothesis
• No Back Tracking– Local minima…
• Inductive bias : approximate “prefer shortest tree”– Information-gain gives a bias for trees with minimal depth
53
Hypothesis Space Search
• Performs batch learning that processes all training instances at once rather than incremental learning that updates a hypothesis after each example.
• Guaranteed to find a tree consistent with any conflict-free training set – i.e. identical feature vectors always assigned the same class,
54
Occam’s Razor
• Occam’s Razor: Prefer the simplest hypothesis that fits the data
William of Ockham (AD 1285? – 1347?)
55
Complex DT : Simple DT
temperature
cool mild hot
outlook outlook outlook
sunny o’cast rain
P N windy
true false
N P
sunny
o’cast rain
windy P humid
true false
P N
high normal
windy P
true false
N P
true false
N humid
high normal
Poutlook
sunny o’cast rain
N P null
outlook
sunny overcast rain
humidity P windy
high normal
N P
true false
N P
56
Inductive Bias in Decision Tree Learning
Occam’s Razor
57
Decision Tree Representation
Decision Trees
58
Disadvantages
• Only allow 2 classes, – This limitation is usually removed in most later
systems.
• Not guaranteed to find the simplest tree.• Not incremental.
– Additional training data can not be considered without rebuilt the the whole tree with all the former data.
59
Advantages
• A reasonably good decision tree without much computation.
• Iterative method usually is found more quickly to build a tree than build on the whole training set.
• ID3 is linear to the difficulty of the problem
60
C4.5 History
• ID3, CHAID – 1960s• C4.5 innovations (Quinlan):
– permit numeric attributes– deal sensibly with missing values– pruning to deal with for noisy data
• C4.5 - one of best-known and most widely-used learning algorithms– Last research version: C4.8, implemented in Weka as
J4.8 (Java)– Commercial successor: C5.0 (available from
Rulequest)
61
C4.5
• ID3 favors attributes with large number of divisions– Lead to overfitting
• Improved version of ID3:– Missing Data– Continuous Data– Pruning
• Subtree replacement by leaf node• Subtree raising
– Automated rule generation– GainRatio: take into account the cardinality of each attribute
values
62
Weakness of ID3: Highly-branching attributes
• Problematic: attributes with a large number of values (extreme case: ID code)• Subsets are more likely to be pure if there is a large number of
values
⇒ Information gain is biased towards choosing attributes with a large number of values
⇒ This may result in overfitting (selection of an attribute that is non-optimal for prediction)
63
Weakness of ID3: Split for ID Code Attribute
• Entropy of split = 0 (since each leaf node is “pure”, having only one case.
• Information gain is maximal for ID code
Day Outlook 온도 Humidity Wind PlayTennis
D1 Sunny Hot High Weak No
D2 Sunny Hot High Strong No
D3 Overcast Hot High Weak Yes
D4 Rain Mild High Weak Yes
D5 Rain Cool Normal Weak Yes
D6 Rain Cool Normal Strong No
D7 Overcast Cool Normal Strong Yes
D8 Sunny Mild High Weak No
D9 Sunny Cool Normal Weak Yes
D10 Rain Mild Normal Weak Yes
D11 Sunny Mild Normal Strong Yes
D12 Overcast Mild High Strong Yes
D13 Overcast Hot Normal Weak Yes
D14 Rain Mild High Strong No
ID code
No Yes NoNo Yes
D1 D2 D3 … D13 D14
64
C4.5• Gain Ratio:
Parent Node, p is split into k partitions
ni is the number of records in partition i
– Adjusts Information Gain by the entropy of the partitioning (SplitINFO). Higher entropy partitioning (large number of small partitions) is penalized!
– Used in C4.5– Designed to overcome the disadvantage of Information Gain– Split information is sensitive to how broadly and uniformly the
attribute splits the data
SplitINFO
GAINGainRATIO Split
split
k
i
ii
nn
nn
SplitINFO1
log
65
Numeric attributes
• Standard method: binary splits– E.g. temp < 45
• Unlike nominal(or, categorical) attributes,every attribute has many possible split points
• Solution is straightforward extension: – Evaluate info gain (or other measure)
for every possible split point of attribute– Choose “best” split point– Info gain for best split point is info gain for attribute
• Computationally more demanding
witten & eibe
66
Example
• Split on temperature attribute:
– E.g. temperature 71.5: yes/4, no/2temperature 71.5: yes/5, no/3
– Info([4,2],[5,3])= 6/14 info([4,2]) + 8/14 info([5,3]) = 0.939 bits
• Place split points halfway between values• Can evaluate all split points in one pass!
64 65 68 69 70 71 72 72 75 75 80 81 83 85
Yes No Yes Yes Yes No No Yes Yes Yes No Yes Yes No
witten & eibe
67
Avoid repeated sorting!
• Sort instances by the values of the numeric attribute– Time complexity for sorting: O (n log n)
• Q. Does this have to be repeated at each node of the tree?
• A: No! Sort order for children can be derived from sort order for parent– Time complexity of derivation: O (n)
– Drawback: need to create and store an array of sorted indices
for each numeric attribute
witten & eibe
68
Weather data – nominal values
Outlook Temperature Humidity Windy Play
Sunny Hot High False No
Sunny Hot High True No
Overcast Hot High False Yes
Rainy Mild Normal False Yes
… … … … …
If outlook = sunny and humidity = high then play = no
If outlook = rainy and windy = true then play = no
If outlook = overcast then play = yes
If humidity = normal then play = yes
If none of the above then play = yes
witten & eibe
69
More speeding up
• Entropy only needs to be evaluated between points of different classes (Fayyad & Irani, 1992)
64 65 68 69 70 71 72 72 75 75 80 81 83 85
Yes No Yes Yes Yes No No Yes Yes Yes No Yes Yes No
Potential optimal breakpoints
Breakpoints between values of the same class cannotbe optimal
valueclass
X
70
Continuous Attributes: Computing Gini Index...
• For efficient computation: for each attribute,– Sort the attribute on values– Linearly scan these values, each time updating the count matrix and
computing gini index– Choose the split position that has the least gini index
Cheat No No No Yes Yes Yes No No No No
Taxable Income
60 70 75 85 90 95 100 120 125 220
55 65 72 80 87 92 97 110 122 172 230
<= > <= > <= > <= > <= > <= > <= > <= > <= > <= > <= >
Yes 0 3 0 3 0 3 0 3 1 2 2 1 3 0 3 0 3 0 3 0 3 0
No 0 7 1 6 2 5 3 4 3 4 3 4 3 4 4 3 5 2 6 1 7 0
Gini 0.420 0.400 0.375 0.343 0.417 0.400 0.300 0.343 0.375 0.400 0.420
Split Positions
Sorted Values
71
Splitting Based on Nominal Attributes
• Multi-way split: Use as many partitions as distinct values.
• Binary split: Divides values into two subsets. Need to find optimal partitioning.
CarTypeFamily
Sports
Luxury
CarType{Family, Luxury} {Sports}
CarType{Sports, Luxury} {Family} OR
Size{Small, Large} {Medium}
72
Splitting Based on Continuous Attributes
TaxableIncome> 80K?
Yes No
TaxableIncome?
(i) Binary split (ii) Multi-way split
< 10K
[10K,25K) [25K,50K) [50K,80K)
> 80K
73
Missing as a separate value
• Missing value denoted “?” in C4.X• Simple idea: treat missing as a separate value• Q: When this is not appropriate?• A: When values are missing due to different
reasons – Example: field IsPregnant=missing for a male patient
should be treated differently (no) than for a female patient of age 25 (unknown)
74
Handling Missing Attribute Values
• Missing values affect decision tree construction in three different ways:– Affects how impurity measures are computed– Affects how to distribute instance with missing
value to child nodes– Affects how a test instance with missing value
is classified
75
Computing Impurity Measure
Tid Refund Marital Status
Taxable Income Class
1 Yes Single 125K No
2 No Married 100K No
3 No Single 70K No
4 Yes Married 120K No
5 No Divorced 95K Yes
6 No Married 60K No
7 Yes Divorced 220K No
8 No Single 85K Yes
9 No Married 75K No
10 ? Single 90K Yes 10
Class = Yes
Class = No
Refund=Yes 0 3
Refund=No 2 4
Refund=? 1 0
Split on Refund:
Entropy(Refund=Yes)
-(0)log(0/3) – (3/3)log(3/3) = 0
Entropy(Refund=No) = -(2/6)log(2/6) – (4/6)log(4/6) = 0.9183
Entropy(Children) = 3/10 (0) + 6/10 (0.9183) = 0.551
Gain = 0.8813 – 0.551 = 0.3303
Missing value
Before Splitting: Entropy(Parent) = -0.3 log(0.3)-(0.7)log(0.7) = 0.8813
Refund
yes no
76
Handling Missing Attribute Values
• Missing values affect decision tree construction in three different ways:– Affects how impurity measures are computed– Affects how to distribute instance with missing value to
child nodes– Affects how a test instance with missing value is
classified
77
Distribute InstancesTid Refund Marital
Status Taxable Income Class
1 Yes Single 125K No
2 No Married 100K No
3 No Single 70K No
4 Yes Married 120K No
5 No Divorced 95K Yes
6 No Married 60K No
7 Yes Divorced 220K No
8 No Single 85K Yes
9 No Married 75K No 10
RefundYes No
Class=Yes 0
Class=No 3
Cheat=Yes 2
Cheat=No 4
RefundYes
Tid Refund Marital Status
Taxable Income Class
10 ? Single 90K Yes 10
No
Class=Yes 2 + 6/ 9
Class=No 4
Probability that Refund=Yes is 3/9
Probability that Refund=No is 6/9
Assign record to the left child with weight = 3/9 and to the right child with weight = 6/9
Class=Yes 0 + 3/ 9
Class=No 3
78
Handling Missing Attribute Values
• Missing values affect decision tree construction in three different ways:– Affects how impurity measures are computed– Affects how to distribute instance with missing value to
child nodes– Affects how a test instance with missing value is
classified
79
Classify Instances
Refund
MarSt
TaxInc
YESNO
NO
NO
YesNo
Married Single,
Divorced
< 80K > 80K
Married Single Divorced Total
Class=No 3 1 0 4
Class=Yes 6/9 1 1 2.67
Total 3.67 2 1 6.67
Tid Refund Marital Status
Taxable Income Class
11 No ? 85K ? 10
New record:
Probability that Marital Status = Married is 3.67/6.67
Probability that Marital Status ={Single,Divorced} is 3/6.67
80
From trees to rules – how?
• How can we produce a set of rules from a decision tree?
81
From trees to rules – simple• Simple way: one rule for each leaf• C4.5rules: greedily prune conditions from each rule
if this reduces its estimated error– Can produce duplicate rules– Check for this at the end
• Then– look at each class in turn– consider the rules for that class– find a “good” subset (guided by MDL)
• Then rank the subsets to avoid conflicts• Finally, remove rules (greedily) if this decreases
error on the training data
witten & eibe
82
C4.5rules: choices and options
• C4.5rules slow for large and noisy datasets• Commercial version C5.0rules uses a different
technique– Much faster and a bit more accurate
witten & eibe
83
CART Split Selection Method
Motivation: We need a way to choose quantitatively between different splitting predicates– Idea: Quantify the impurity of a node– Method: Select splitting predicate that
generates children nodes with minimum impurity from a space of possible splitting predicates
84
CART
• If a data set D contains examples from n classes, gini index, gini(D) is defined as
where pj is the relative frequency of class j in D
• If a data set D is split on A into two subsets D1 and D2, the gini
index gini(D) is defined as
• Reduction in Impurity:
• The attribute provides the smallest ginisplit(D) (or the largest
reduction in impurity) is chosen to split the node
n
jp jDgini
1
21)(
)(||||)(
||||)( 2
21
1 DginiDD
DginiDDDginiA
)()()( DginiDginiAginiA
85
Measure of Impurity: GINI
– Maximum (0.5) when records are equally distributed among all classes, implying least interesting information
– Minimum (0.0) when all records belong to one class, implying most interesting information
C1 0C2 6
Gini=0.000
C1 2C2 4
Gini=0.444
C1 3C2 3
Gini=0.500
C1 1C2 5
Gini=0.278
n
jp jDgini
1
21)(
86
Examples for computing GINI
C1 0 C2 6
C1 2 C2 4
C1 1 C2 5
P(C1) = 0/6 = 0 P(C2) = 6/6 = 1
Gini = 1 – P(C1)2 – P(C2)2 = 1 – 0 – 1 = 0
P(C1) = 1/6 P(C2) = 5/6
Gini = 1 – (1/6)2 – (5/6)2 = 0.278
P(C1) = 2/6 P(C2) = 4/6
Gini = 1 – (2/6)2 – (4/6)2 = 0.444
n
jp jDgini
1
21)(
87
Comparison among Splitting Criteria
For a 2-class problem:
88
CART
• Ex. D has 9 tuples in buys_computer = “yes” and 5 in “no”
• Suppose the attribute income partitions D into 10 in D1: {low, medium}
and 4 in D2
but gini{medium,high} is 0.30 and thus the best since it is the lowest
• All attributes are assumed continuous-valued
• Can be modified for categorical attributes
459.014
5
14
91)(
22
Dgini
)(14
4)(
14
10)( 11},{ DGiniDGiniDgini mediumlowincome
89
Comparing Attribute Selection Measures
• The three measures, in general, return good results but– Information gain:
• biased towards multivalued attributes
– Gain ratio: • tends to prefer unbalanced splits in which one partition is
much smaller than the others
– Gini index: • biased to multivalued attributes• has difficulty when # of classes is large• tends to favor tests that result in equal-sized partitions
and purity in both partitions
90
Issues in Decision Tree Learning
Overfitting in Decision Trees
Underfitting and Overfitting
Overfitting
Underfitting: when model is too simple, both training and test errors are large
91
92
Issues in Decision Tree Learning
Overfitting in Decision Treesh
h’
Overfitting due to Noise
Decision boundary is distorted by noise point
93
94
Overfitting
• Learning a tree that classifies the training data perfectly may not lead to the tree with the best generalization to unseen data.– There may be noise in the training data that the tree is
erroneously fitting.– The algorithm may be making poor decisions towards the
leaves of the tree that are based on very little data and may not reflect reliable trends.
hypothesis complexity
accu
racy
on training data
on test data
95
Overfitting Example
voltage (V)
curr
ent (
I)
Testing Ohms Law: V = IR (I = (1/R)V)
Ohm was wrong, we have found a more accurate function!
Perfect fit to training data with an 9th degree polynomial(can fit n points exactly with an n-1 degree polynomial)
Experimentallymeasure 10 points
Fit a curve to theResulting data.
96
Overfitting Example
voltage (V)
curr
ent (
I)
Testing Ohms Law: V = IR (I = (1/R)V)
Better generalization with a linear functionthat fits training data less accurately.
97
Overfitting Noise in Decision Trees
• Category or feature noise can easily cause overfitting.– Add noisy instance <medium, blue, circle>: pos (but really neg)
shape
circle square triangle
color
red bluegreen
pos neg pos
neg neg
98
Overfitting Noise in Decision Trees
• Category or feature noise can easily cause overfitting.– Add noisy instance <medium, blue, circle>: pos (but really neg)
shape
circle square triangle
color
red bluegreen
pos neg pos
neg
<medium, blue, circle>: +<big, blue, circle>:
small med big
posneg neg
99
Overfitting Prevention (Pruning) Methods
• Two basic approaches for decision trees– Prepruning: Halt tree construction early—do not split a node
if this would result in the goodness measure falling below a thresholdDifficult to choose an appropriate threshold
– Postpruning: Remove branches from a “fully grown” tree —get a sequence of progressively pruned treesUse a set of data different from the training data to
decide which is the “best pruned tree”
100
Overfitting Prevention (Pruning) Methods
• Label leaf resulting from pruning with the majority class of the remaining data, or a class probability distribution.
• Method for determining which subtrees to prune:– Cross-validation: Reserve some training data as a hold-out set
(validation set) to evaluate utility of subtrees.– Statistical test: Use a statistical test on the training data to
determine if any observed regularity can be dismissed as likely due to random chance.
– Minimum description length (MDL): Determine if the additional complexity of the hypothesis is less complex than just explicitly remembering any exceptions resulting from pruning.
Minimum Description Length (MDL)
• Cost(Model,Data) = Cost(Data|Model) + Cost(Model)– Cost is the number of bits needed for encoding.– Search for the least costly model.
• Cost(Data|Model) encodes the misclassification errors.• Cost(Model) = node encoding(number of children) + splitting condition encoding
A B
A?
B?
C?
10
0
1
Yes No
B1 B2
C1 C2
X yX1 1X2 0X3 0X4 1
… …Xn 1
X yX1 ?X2 ?X3 ?X4 ?
… …Xn ?
101
102
Pruning• Goal: Prevent overfitting to noise in the data• Two strategies for “pruning” the decision tree:
– (Stop earlier / Forward pruning): Stop growing the tree earlier – extra stopping conditions, e.g.• Stop if all instances belong to the same class• Stop if all the attribute values are the same• Stop if number of instances < some user-specified threshold• Stop if expanding the current node does not improve impurity
measures (e.g., Gini or Gain).
– (Post-pruning): Allow overfit and then post-prune the tree.• Estimation of errors and tree size to decide which subtree should be pruned.
• Postpruning preferred in practice—prepruning can “stop too early”
103
Early stopping
• Pre-pruning may stop the growth process prematurely: early stopping
• But: XOR-type problems rare in practice
• And: pre-pruning faster than post-pruning
witten & eibe
104
Post-pruning
• First, build full tree• Then, prune it
– Fully-grown tree shows all attribute interactions
• Two pruning operations: – Subtree replacement– Subtree raising
• Possible strategies:– error estimation– significance testing– MDL principle
witten & eibe
105
Post-pruning: Subtree replacement, 1
• Bottom-up• Consider replacing a tree
only after considering all its subtrees
• Ex: labor negotiations
witten & eibe
106
Post-pruning: Subtree replacement, 2
What subtree can we replace?
107
Subtreereplacement, 3
• Bottom-up• Consider replacing a tree
only after considering all its subtrees
witten & eibe
108
*Subtree raising• Delete node
• Redistribute instances
• Slower than subtree replacement
• (Worthwhile?)
witten & eibe
X
109
Estimating error rates
• Prune only if it reduces the estimated error• Error on the training data is NOT a useful estimator
Q: Why it would result in very little pruning?• Use hold-out set for pruning
(“reduced-error pruning”)• C4.5’s method
– Derive confidence interval from training data– Use a heuristic limit, derived from this, for pruning– Standard Bernoulli-process-based method– Shaky statistical assumptions (based on training data)
witten & eibe
Estimating error rates
• Binomial확률 변수 성공 확률이 p 인 베르누이 시행을 n 번 독립적으로 반복했을 때의 성공 횟수 , X independently and identically distribution (iid) 이항 분포 X 의 확률 분포를 시행횟수 n, 성공률 p 를 갖는 이항 분포라 한다 .
확률 질량 함수 평균 E(X)=np
분산 Var(X)=np(1-p)
nxppx
nxXP xnx ,,2,1,0,)1()(
0419.0
)1(5.020
50)20( 205020
pXP
50 번 시행 중 Head 가 20 번 나올 확률 ?
110
111
*Mean and variance
• Mean and variance for a Bernoulli trial: p, p (1–p)
• Expected error(or success) rate f=S/N
• Mean and variance for f : p, p (1–p)/N
• For large enough N, f follows a Normal distribution
• c% confidence interval [–z T z] for random variable with
0 mean is given by:
• With a symmetric distribution:
czTz ]Pr[
]Pr[21]Pr[ zTzTz
witten & eibe
112
*Confidence limits• Confidence limits for the normal distribution with 0 mean and a
variance of 1:
• Thus:
• To use this we have to reduce our random variable f to have 0 mean and unit variance
Pr[X z] z
0.1% 3.09
0.5% 2.58
1% 2.33
5% 1.65
10% 1.28
20% 0.84
25% 0.69
40% 0.25
%90
]65.1Pr[*21]65.165.1Pr[
XX
–1 0 1 1.65
Confidence Interval on the Mean of aNormal Distribution, Variance Known
113
Npppf
/)1(
Mean p and variance p (1–p)/N for f(error rate) of sample
We may standardize f by f – p and divide by Npp /)1(
114
*Transforming f
• Transformed random variable for f :
(i.e. subtract the mean and divide by the standard deviation)
• Resulting equation of confidence limits:
• Solving for mean error rate p from a normal distribution for f :
Npppf
/)1(
czNpp
pfz
/)1(Pr
Nz
Nz
Nf
Nf
zN
zfp
2
2
222
142
115
C4.5’s method
• Error estimate for subtree is weighted sum of error estimates for all its leaves
• Error estimate for a node (upper bound):
• If c = 25% then z = 0.69 (from normal distribution)• f is the error on the training data• N is the number of instances covered by the leaf
witten & eibe
Nz
Nz
Nf
Nf
zN
zfp
2
2
222
142
116
Example
f=0.33 p=0.47
f=0.5 p=0.72
f=0.33 p=0.47
f = 5/14 p = 0.46p < 0.51so prune!
Combined using ratios 6:2:6 e=0.51witten & eibe
Example
117
Issues in Decision Tree Learning
Effect of Reduced-Error Pruning
118
Issues in Decision Tree Learning
Rule Post-Pruning
119
Issues in Decision Tree Learning
Converting A Tree to Rules
Model Evaluation
• Metrics for Performance Evaluation– How to evaluate the performance of a model?
• Methods for Performance Evaluation– How to obtain reliable estimates?
• Methods for Model Comparison– How to compare the relative performance among competing
models?
120
Metrics for Performance Evaluation
• Focus on the predictive capability of a model– Rather than how fast it takes to classify or build models,
scalability, etc.
• Confusion Matrix:
PREDICTED CLASS
ACTUALCLASS
Class=Yes Class=No
Class=Yes a b
Class=No c d
a: TP (true positive)
b: FN (false negative)
c: FP (false positive)
d: TN (true negative)
121
Metrics for Performance Evaluation…
• Most widely-used metric:
PREDICTED CLASS
ACTUALCLASS
Class=Yes Class=No
Class=Yes a(TP)
b(FN)
Class=No c(FP)
d(TN)
FNFPTNTPTNTP
dcbada
Accuracy
122
Limitation of Accuracy
• Consider a 2-class problem– Number of Class 0 examples = 9990– Number of Class 1 examples = 10
• If model predicts everything to be class 0, accuracy is 9990/10000 = 99.9 %– Accuracy is misleading because model does not
detect any class 1 example
123
Cost Matrix
PREDICTED CLASS
ACTUALCLASS
C(i|j) Class=Yes Class=No
Class=Yes C(Yes|Yes) C(No|Yes)
Class=No C(Yes|No) C(No|No)
C(i|j): Cost of misclassifying class j example as class i
124
Computing Cost of Classification
Cost Matrix
PREDICTED CLASS
ACTUALCLASS
C(i|j) + -
+ -1 100
- 1 0
Model M1 PREDICTED CLASS
ACTUALCLASS
+ -
+ 150 40
- 60 250
Model M2 PREDICTED CLASS
ACTUALCLASS
+ -
+ 250 45
- 5 200
Accuracy = 80%
Cost = 3910
Accuracy = 90%
Cost = 4255
125
Cost vs Accuracy
Count PREDICTED CLASS
ACTUALCLASS
Class=Yes Class=No
Class=Yes a b
Class=No c d
Cost PREDICTED CLASS
ACTUALCLASS
Class=Yes Class=No
Class=Yes p q
Class=No q p
N = a + b + c + d
Accuracy = (a + d)/N
Cost = p (a + d) + q (b + c)
= p (a + d) + q (N – a – d)
= q N – (q – p)(a + d)
= N [q – (q-p) Accuracy]
Accuracy is proportional to cost if1. C(Yes|No)=C(No|Yes) = q 2. C(Yes|Yes)=C(No|No) = p
126
Cost-Sensitive Measures
cbaa
prrp
baa
caa
222
(F) measure-F
(r) Recall
(p)Precision
Precision is biased towards C(Yes|Yes) & C(Yes|No) Recall is biased towards C(Yes|Yes) & C(No|Yes) F-measure is biased towards all except C(No|No)
dwcwbwawdwaw
4321
41Accuracy Weighted
127
Model Evaluation
• Metrics for Performance Evaluation– How to evaluate the performance of a model?
• Methods for Performance Evaluation– How to obtain reliable estimates?
• Methods for Model Comparison– How to compare the relative performance among
competing models?
128
Methods for Performance Evaluation
• How to obtain a reliable estimate of performance?
• Performance of a model may depend on other factors besides the learning algorithm:– Class distribution– Cost of misclassification– Size of training and test sets
129
Learning Curve
Learning curve shows how accuracy changes with varying sample size
Effect of small sample size:
- Bias in the estimate
- Variance of estimate
130
target
Bias
Variance
131
Issues with Reduced Error Pruning
• The problem with this approach is that it potentially “wastes” training data on the validation set.
• Severity of this problem depends where we are on the learning curve:
test
acc
urac
y
number of training examples
Model Evaluation
• Metrics for Performance Evaluation– How to evaluate the performance of a model?
• Methods for Performance Evaluation– How to obtain reliable estimates?
• Methods for Model Comparison– How to compare the relative performance among
competing models?
132
ROC (Receiver Operating Characteristic)
• Developed in 1950s for signal detection theory to analyze noisy signals – Characterize the trade-off between positive hits and false alarms
• ROC curve plots TP (on the y-axis) against FP (on the x-axis)
• Performance of each classifier represented as a point on the ROC curve– changing the threshold of algorithm, sample distribution or cost
matrix changes the location of the point
133
How to Construct an ROC curve
Instance P(+|A) True Class
1 0.95 +
2 0.93 +
3 0.87 -
4 0.85 -
5 0.85 -
6 0.85 +
7 0.76 -
8 0.53 +
9 0.43 -
10 0.25 +
• Use classifier that produces posterior probability for each test instance P(+|A)
• Sort the instances according to P(+|A) in decreasing order
• Apply threshold at each unique value of P(+|A)
• Count the number of TP, FP, TN, FN at each threshold
• TP rate, TPR = TP/(TP+FN)
• FP rate, FPR = FP/(FP + TN)
134
How to construct an ROC curve
Class + - + - - - + - + +
P 0.25 0.43 0.53 0.76 0.85 0.85 0.85 0.87 0.93 0.95 1.00
TP 5 4 4 3 3 3 3 2 2 1 0
FP 5 5 4 4 3 2 1 1 0 0 0
TN 0 0 1 1 2 3 4 4 5 5 5
FN 0 1 1 2 2 2 2 3 3 4 5
TPR 1 0.8 0.8 0.6 0.6 0.6 0.6 0.4 0.4 0.2 0
FPR 1 1 0.8 0.8 0.6 0.4 0.2 0.2 0 0 0
Threshold
ROC Curve: TP rate, TPR = TP/(TP+FN)
FP rate, FPR = FP/(FP + TN)
(1,1)
(1, 0.8)(0.8, 0.8)
P(+|A): A 가 발생 할때의 + 의 조건부 확률
Example
Actual Predicate(t ≤ )
+ 0.43≤0.25 FN
- 0.43≤0.43 FP
+ 0.43≤0.53 TP
- 0.43≤0.76 FP
- 0.43≤0.85 FP
- 0.43≤0.85 FP
+ 0.43≤0.85 TP
- 0.43≤0.87 FP
+ 0.43≤0.93 TP
+ 0.43≤0.95 TP
135
ROC Curve
At threshold t:
TP=0.5, FN=0.5, FP=0.12, TN=0.88
- 1-dimensional data set containing 2 classes (positive and negative)
- any points located at x >= t is classified as positive, P(+| x >= t ):
TP rate, TPR = TP/(TP+FN) = 0.5/ (0.5+0.5) = 0.5
FP rate, FPR = FP/(FP + TN) = 0.5 / (0.12+0.88) = 0.5 136
ROC Curve
(TP,FP):• (0,0): declare everything
to be negative class• (1,1): declare everything
to be positive class• (1,0): ideal
• Diagonal line:– Random guessing– Below diagonal line:
• prediction is opposite of the true class
ideal
everythingto be negative class
everythingto be positive class
137
Using ROC for Model Comparison
No model consistently outperform the other
M1 is better for small FPR
M2 is better for large FPR
Area Under the ROC curve Ideal:
Area = 1 Random guess:
Area = 0.5
138
Test of Significance
• Given two models:– Model M1: accuracy = 85%, tested on 30 instances– Model M2: accuracy = 75%, tested on 5000 instances
• Can we say M1 is better than M2?– How much confidence can we place on accuracy of M1 and M2?– Can the difference in performance measure be explained as a
result of random fluctuations in the test set?
139
Confidence Interval for Accuracy
• Prediction can be regarded as a Bernoulli trial– A Bernoulli trial has 2 possible outcomes
– Possible outcomes for prediction: correct or wrong
– Collection of Bernoulli trials has a Binomial distribution:
• x Bin(N, p) x: number of correct predictions
• e.g: Toss a fair coin 50 times, how many heads would turn up? Expected number of heads = Np = 50 0.5 = 25
• Given x (# of correct predictions) or equivalently, acc=x/N, and N (# of test instances),
Can we predict p (true accuracy of model)?
140
Confidence Interval for Accuracy
• For large test sets (N > 30), – f(success rate) has a normal distribution
with mean p and variance p(1-p)/N
• Confidence Interval for p:
1
)/)1(
( 2/12/ ZNpp
pfZP
Area = 1 -
Z/2 Z1- /2
141
)(2
4422
2/
222/2/
22/
ZN
fNfNZZZfNp
Confidence Interval for Accuracy
• Consider a model that produces an accuracy of 80% when evaluated on 100 test instances:– N=100, acc = 0.8
– Let 1- = 0.95 (95% confidence)
– From probability table, Z/2=1.96 1- Z
0.99 2.58
0.98 2.33
0.95 1.96
0.90 1.65
N 50 100 500 1000 5000
p(lower) 0.670 0.711 0.763 0.774 0.789
p(upper) 0.888 0.866 0.833 0.824 0.811
)(2
4422
2/
222/2/
22/
ZN
accNaccNZZZaccNp
142
Comparing Performance of 2 Algorithms
• Each learning algorithm may produce k models:– L1 may produce M11 , M12, …, M1k– L2 may produce M21 , M22, …, M2k
• If models are generated on the same test sets D1,D2, …, Dk (e.g., via cross-validation)– For each set: compute dj = e1j – e2j
– dj has mean dt and variance t
– Estimate:
tkt
k
jj
t
zdd
kk
dd
ˆ
)1(
)(ˆ
1,1
1
2
2
143
Comparing Performance of 2 Models
• Given two models, say M1 and M2, which is better?– M1 is tested on D1 (size=n1), found error rate = e1
– M2 is tested on D2 (size=n2), found error rate = e2
– Assume D1 and D2 are independent
– If n1 and n2 are sufficiently large, then
– Approximate:
222
111
,~
,~
Ne
Ne
i
ii
i nee )1(
ˆ
144
Comparing Performance of 2 Models
• To test if performance difference is statistically significant: d = e1 – e2
– d ~ NN(dt,t) where dt is the true difference
– Since D1 and D2 are independent, their variance adds up:
– At (1-) confidence level,
2
22
1
11
22
21
22
21
2
)1()1(
ˆˆ
n
ee
n
eet
ttZdd
ˆ
2/
145
An Illustrative Example
• Given: M1: n1 = 30, e1 = 0.15 M2: n2 = 5000, e2 = 0.25
• d = |e2 – e1| = 0.1 (2-sided test)
• At 95% confidence level, Z/2=1.96
Interval contains difference may not be statistically significant
0043.05000
)25.01(25.030
)15.01(15.0ˆ
d
128.0100.00043.096.1100.0 t
d
146
147
Issues in Decision Tree Learning
• Weakness of decision trees
– Not Always sufficient to learn complex concepts (e.g., weighted evaluation function)
– Can be hard to understand. Real problems can produce deep trees with a large branching factor
– Some problems with continuously-valued attributes or classes may not be easily discretized
– Methods for handling missing attribute values are somewhat clumsy
148
Issues in Decision Tree Learning
• Better splitting criteria– Information gain prefers features with many values.
• Continuous features• Predicting a real-valued function (regression trees)• Missing feature values• Features with costs• Misclassification costs• Mining large databases that do not fit in main memory
Issues in Decision Tree Learning
• Data Fragmentation
• Search Strategy
• Expressiveness
• Tree Replication
149
Issues in Decision Tree Learning
• Number of instances gets smaller as you traverse down the tree
• Number of instances at the leaf nodes could be too small to make any statistically significant decision
150
Data Fragmentation
Search Strategy
• Finding an optimal decision tree is NP-hard
• The algorithm presented so far uses a greedy, top-down, recursive partitioning strategy to induce a reasonable solution
• Other strategies?– Bottom-up– Bi-directional
151
Issues in Decision Tree Learning
Expressiveness
• Decision tree provides expressive representation for learning discrete-valued function– But they do not generalize well to certain types of Boolean functions
• Example: parity function: – Class = 1 if there is an even number of Boolean attributes with truth value = True– Class = 0 if there is an odd number of Boolean attributes with truth value = True
• For accurate modeling, must have a complete tree
• Not expressive enough for modeling continuous variables– Particularly when test condition involves only a single attribute at-a-time
152
Issues in Decision Tree Learning
Decision Boundary
y < 0.33?
: 0 : 3
: 4 : 0
y < 0.47?
: 4 : 0
: 0 : 4
x < 0.43?
Yes
Yes
No
No Yes No
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
y
• Border line between two neighboring regions of different classes is known as decision boundary
• Decision boundary is parallel to axes because test condition involves a single attribute at-a-time
153
Issues in Decision Tree Learning
Oblique Decision Trees
x + y < 1
Class = + Class =
• Test condition may involve multiple attributes
• More expressive representation
• Finding optimal test condition is computationally expensive
154
Issues in Decision Tree Learning
Tree Replication
P
Q R
S 0 1
0 1
Q
S 0
0 1
• Same subtree appears in multiple branches
155
Issues in Decision Tree Learning