Basics of Decision Trees

A flow-chart-like hierarchical tree structure– Often restricted to a binary structure

Root: represents the entire datasetA node without children is called a leaf node.

Otherwise is called an internal node.– Internal nodes: denote a test on an attribute– Branch (split): represents an outcome of the test

• Univariate split (based on a single attribute)• Multivariate split

– Leaf nodes: represent class labels or class distribution

Example (training dataset)

Three predictor attributes: salary, age, employment Class label attribute: group

Record ID Salary Age Employment Group

1 30K 30 Self C

2 40K 35 Industry C

3 70K 50 Academia C

4 60K 45 Self B

5 70K 30 Academia B

6 60K 35 Industry A

7 60K 35 Self A

8 70K 30 Self A

9 40K 45 Industry C

Example (univariate split)

Each internal node of a decision tree is labeled with a predictor attribute, called the splitting attribute

each leaf node is labeled with a class label.

Each edge originating from an internal node is labeled with a splitting predicate that involves only the node’s splitting attribute.

The combined information about splitting attributes and splitting predicates at a node is called the split criterion

Salary

Age

Employment

Group C

Group C

Group B Group A

<= 50K > 50K

<= 40 > 40

Academia, Industry Self

Two Phases

Most decision tree generation consists of two phases– Tree construction– Tree pruning

Building Decision Trees

Basic algorithm (a greedy algorithm)– Tree is constructed in a top-down recursive divide-

and-conquer manner– At start, all the training examples are at the root– Attributes are categorical (if continuous-valued, they

are discritized in advance)– Examples are partitioned recursively based on

selected attributes– Test attributes are selected on the basis of a heuristic

or statistical measure (e.g., information gain)

Tree-Building Algorithm

We can build the whole tree by calling:

BuildTree(dataset TrainingData, split-selection-method CL)

Input: dataset S, split-selection-method CLOutput: decision tree for S

Top-Down Decision Tree Induction Schema (Binary Splits):BuildTree(dataset S, split-selection-method CL)(1) If (all points in S are in the same class) then return;(2) Using CL to evaluate splits for each attribute(3) Use best split found to partition S into S1 and S2;(4) BuildTree(S1, CL)(5) BuildTree(S2, CL)

Split Selection

Information gain / Gain ratio (ID3/C4.5)– All attributes are assumed to be categorical– Can be modified for continuous-valued attributes

Gini index (IBM IntelligentMiner)– All attributes are assumed continuous-valued– Assume there exist several possible split values for

each attribute– May need other tools, such as clustering, to get the

possible split values– Can be modified for categorical attributes

Information Gain (ID3) T – Training set; S – any set of cases

freq(Ci, S) – the number of cases that belong to class Ci

|S| -- the number of cases in set S Information of set S is defined:

consider a similar measurement after T has been partitioned in accordance with the n outcomes of a test X. The expected information requirement can be found as the weighted sum over the subsets, as

The quantity

measures the information that is gained by partitioning T in accordance with the test X. The gain criterion, then, selects a test to maximize this information gain.

||

),(log

||

),()( 2

1 S

SCfreq

S

SCfreqSinfo j

k

j

j

i

n

i

iX Tinfo

T

TTinfo

1 ||

||)(

gain(X) = info(T) – infoX(T)

Gain Ratio (C4.5)

Gain criterion has a serious deficiency – it has a strong bias in favor of tests with many outcomes.

We have

This represents the potential information generated by dividing T into n subsets, whereas the information gain measures the information relevant to classification that arises from the same division.

Then,

expresses the proportion of information generated by the split that is useful, i.e., that appears helpful for classification.

n

i

ii

T

T

T

TXinfo split

12 ||

||log

||

||

Gain ratio(X) = gain(X)/split info(X)

Gini Index (IBM IntelligentMiner)

If a data set T contains examples from n classes, gini index, gini(T) is defined as

where pj is the relative frequency of class j in T. If a data set T is split into two subsets T1 and T2 with sizes N1 and

N2 respectively, the gini index of the split data contains examples from n classes, the gini index gini(T) is defined as

The attribute provides the smallest ginisplit(T) is chosen to split the node (need to enumerate all possible splitting points for each attribute).

n

jp jTgini

1

21)(

)()()( 22

11 Tgini

NN

TginiNNTginisplit

Pruning Decision Trees

Why prune?– Overfitting

• Too many branches, some may reflect anomalies due to noise or outliers

• Result is in poor accuracy for unseen samples Two approaches to prune

– Prepruning: Halt tree construction early—do not split a node if this would result in the goodness measure falling below a threshold

• Difficult to choose an appropriate threshold– Postpruning: Remove branches from a “fully grown” tree—get a

sequence of progressively pruned trees• Use a set of data different from the training data to decide

which is the “best pruned tree”

Well-known Pruning Methods

Reduced Error PruningPessimistic Error PruningMinimum Error PruningCritical Value PruningCost-Complexity PruningError-Based Pruning

Extracting Rules from Decision Trees

Represent the knowledge in the form of IF-THEN rules One rule is created for each path from the root to a leaf Each attribute-value pair along a path forms a

conjunction The leaf node holds the class prediction Rules are easier for humans to understand Example

– IF salary=“>50k” AND age=“>40” THEN group=“C”

Why Decision Tree Induction

Compared to a neural network or a Bayesian classifier, a decision tree is easily interpreted/comprehended by humans

While training neural networks can take large amounts of time and thousands of iterations, inducing decision trees is efficient and is thus suitable for large training sets

Decision tree generation algorithms do not require additional information besides that already contained in the training data

Decision trees display good classification accuracy compared to other techniques

Decision tree induction can use SQL queries for accessing databases

Tree Quality Measures

AccuracyComplexity

– Tree size– Number of leaf nodes

Computational speed

Scalability Scalability: Classifying data sets with millions of examples and

hundreds of attributes with reasonable speed 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

PUBLIC (VLDB’98 — Rastogi & Shim)– integrates tree splitting and tree pruning: stop growing the tree earlier

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 CLOUDS

Other Issues

Allow for continuous-valued attributes– Dynamically define new discrete-valued attributes that partition the

continuous attribute value into a discrete set of intervals

Handle missing attribute values– Assign the most common value of the attribute

– Assign probability to each of the possible values

Attribute (feature) construction– Create new attributes based on existing ones that are sparsely represented

– This reduces fragmentation, repetition, and replication

Incremental tree induction Integration of data warehousing techniques Different data access methods Bias in split selection

Decision Tree Induction Using P-trees

Basic Ideas– Calculate information gain, gain ratio or gini index by

using the count information recorded in P-trees.– P-tree generation replaces sub-sample set creation.– Use P-tree to determine if all the samples are in the

same class. – Without additional database scan

Using P-trees to CalculateInformation Gain/Gain Ratio

C – class label attributePs – P-tree of set SFreq(Cj, S) = rc{Ps ^ Pc(Vcj)}|S| = rc{Ps}|Ti| = rc{PT^P(VXi)}|T| = rc{PT}So every formula of Information Gain and Gain

Ratio can be calculated directly using P-trees.

P-Classifier versus ID3

Classification cost with respect to the dataset size

Classification Time

0

100

200

300

400

500

600

700

0 20 40 60 80

Size of data (M)

ID3

P-Classifer