Scribd
Upload
228 views

Decision Tree

This document discusses segmentation using decision trees. It provides an overview of the SAS Data Mining Solution, introduces decision trees, and covers assessment measures. Decision trees allow businesses to define segments based on rules derived from data. The SAS DataSplits tool implements both CHAID and CART decision tree methodologies within SAS Enterprise Miner for segmentation, profiling clusters, classification modeling, and variable selection/creation. Model assessment considers metrics like classification, sensitivity, lift, and profit/loss equations.

Uploaded by

Sarbani Dasgupts
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
228 views

Decision Tree

This document discusses segmentation using decision trees. It provides an overview of the SAS Data Mining Solution, introduces decision trees, and covers assessment measures. Decision trees allow businesses to define segments based on rules derived from data. The SAS DataSplits tool implements both CHAID and CART decision tree methodologies within SAS Enterprise Miner for segmentation, profiling clusters, classification modeling, and variable selection/creation. Model assessment considers metrics like classification, sensitivity, lift, and profit/loss equations.

Uploaded by

Sarbani Dasgupts
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 52

Segmentation using

Decision Trees
Gerhard Held
Product Manager Analytical Applications

SAS Institute Europe

Outline
1. The SAS Data Mining Solution
2. Introduction to Decision Trees
3. Assessment Measures
4. Overview of DataSplits Tool within SAS
Data Mining Solution

Data Mining Solutions?


??? ??? ??? ??? ???
???
???
???
???
???

???
???
???
???
???
??? ??? ??? ??? ???

The SAS Data Mining


Solution
Data Mining is a Process
Data Mining involves close Co-operation of
IT, Business, and Data Miners

The SAS Data Mining


Solution

Business Problem

The SAS Data Mining


Solution

Business Problem

The SAS Data Mining


Solution
The SAS Data Mining Solution is unique:
IT: DW Access, Scalability
Business Users:
Intuitive Interface and
Business Orientation
Data Miners: Analytical Depth
and Flexibility

Introduction to Decision
Trees
Searching for Segments
What is a Decision Tree?
Common Applications
Tree Methodologies

Searching for Segments

All Observations
4,000 Resp - 29%
10,000 NonResp - 71%
Split on AGE

TEENS <20
3,000 Resp - 60%
2,000 NonResp - 40%

ADULTS >=20
1,000 Resp - 11%
8,000 NonResp - 89%

Searching for Segments


Total Response %
100%
80%
60%
40%
20%
0%
No Reply

Bought

Search
Searchfor
forpredictor
predictorwith
withlarge
largedifference
difference
in
inresponse
responserate
rateover
overits
itscategories.
categories.

Searching for Segments


Total Response %
100%
80%
60%
40%
20%
0%
No Reply

Bought

100%
80%
60%

Young: No or Yes?

40%
20%
0%
No Reply
No

Bought
No

No Reply
Yes

Bought
Yes

Searching for Segments


100%
80%
60%

Young: No or Yes?

40%
20%
0%
No Reply Bought
No
No

No Reply Bought
Yes
Yes

100%
80%
60%
40%

UK: No or Yes?

20%
0%
No Repy Bought
No
No

No Reply Bought
Yes
Yes

Searching for Segments


All Observations
All Observations
4,000 Resp - 29%
10,000 NonResp - 71%
Split on AGE

Searching for Segments

All Observations
4,000 Resp - 29%
10,000 NonResp - 71%
Split on AGE

TEENS <20
3,000 Resp - 60%
2,000 NonResp - 40%

ADULTS >=20
1,000 Resp - 11%
8,000 NonResp - 89%

Searching for Segments


All Observations
4,000 Resp - 29%
10,000 NonResp - 71%
Split on AGE

TEENS <20

ADULTS >=20

3,000 Resp - 60%


2,000 NonResp - 40%

1,000 Resp - 11%


8,000 NonResp - 89%

-1-

Split on Country

UK only

Continent

0 Resp - 0%
1,000 NonResp - 100%

1,000 Resp - 12.5%


7,000 NonResp - 87.5%

- 2-

-3-

Rules Define the Segments


If the person is a teenager
If the person is an adult in the UK ...
Otherwise

IF AGE LT 20 THEN SEGMENT=1;


ELSE IF AGE GE 20 AND CTRY EQ UK
THEN SEGMENT=2;
ELSE SEGMENT=3;

What is a Decision Tree?


A Set of Rules
For defining Data Segments
with similar Values of a Target

Why Trees are Popular?


Easy to Understand
Segmentation defined without Formulae
Express Rules using a single Variable
User can create and modify Rules
Easy to Use
Accepts ordered and unordered Variables
No transformation of Predictors
Uses missing Data

Decision Tree Applications


Segmentation Using a Target Variable
Profiling Clusters
Modeling for Classification
Variable Selection and Creation
A few Examples...

Profiling Clusters - Why...


Group Customers with common
Preferences
No Need to define Target Variable
without a specific Product in Mind
Select Products or Promotions separately

Profiling Clusters - Steps...


Cluster: Form Segments from Data
Profile: Describe Segments intuitively

Clusters & Profiles


Segment 1
ages: 18-25
income: 40K

Startups
Foreign Economical Cars
Junk Food

Segment 2
ages: 26-35
income: 100K

DINKS
Sports Cars
Fashionable Restaurants

Profiling Using Trees


Cluster on demographic Data
Such as: Age, Income, City
Mostly Available, generic
Assign an ID to each Segment
Train a Decision Tree
Segment ID as Target
Profiling Data as Input Variables

New Applications:
Variable Selection
Include
AGE
COUNTRY

AGE

Exclude
COUNTRY

FINANCIAL ASSETS
INCOME

Variable Creation
Include
AGE
STATE

AGE

Create
COUNTRY

TEENS
ADULTS
ADULTS not UK
ADULTS in UK

Tree Methodologies The AID Family


AID (1964), CHAID (1978), FIRM, TREEDISC,
KnowledgeSeeker
Automatic Interaction Detection
Detects complex statistical Relationships
Number of Branches varies from 2 to number of Categories
of Predictor
Statistical Significance Tests used to determine size of Tree
Nominal Responses
Nominal or Ordinal Predictors

Tree Methodologies
Methodologies -Tree
The CART
CART Family
Family
The
CART, Salford Systems, S-Plus (all 1980s)

Classification And Regression Trees


Statistical Prediction
Exactly 2 Branches from each nonterminal Node
Cross-validation and pruning used to determine Size of Tree
Quantitative or nominal Response Variable
is x(1)<5.0?
Nominal, ordinal and continuous Predictors
yes

no

is x(3)<3.2?
yes

class 2

no

class 1

is x(1)<7.6?
yes

class 3

no

class 3

Model Assessment
Classification
Sensitivity
Lift
Profit and Loss Equations

Model Assessment
Alternatives
Proportion correctly predicted

(classified)
Proportion responding to Promotion
Expected Profit

Classification
PREDICTED
O
B
S
E
R
V
E
D

Default
Default
Paid

Paid

10

10

20

40

40

10

50

60

( 10 + 40) / 60 = 50 / 60

Sensitivity
PREDICTED
O
B
S
E
R
V
E
D

Default

Paid

10

10

20

40

40

10

50

60

Default
Paid

Sensitivity: 10 / 20

Sensitivity and
Classification
PREDICTED
O
B
S
E
R
V
E
D

Default

Paid

Default

20

20

Paid

20

20

40

40

20

60

Sensitivity: 20 / 20

( 20 + 20) / 60 = 40 / 60

Specificity
PREDICTED
O
B
S
E
R
V
E
D

Default

Paid

Default

20

20

Paid

20

20

40

40

20

60

Specificity = 20 / 40

The Alternative
of Sensitivity

Lift
PREDICTED
O
B
S
E
R
V
E
D

Respond

No

Respond

20

20

No

20

20

40

40

20

60

Lift = (20/40) / (20/60)

Profit & Loss Equation


Given: A scored Data Set
A Cost Function
Then the Assessment Tool produces for each Level
of the Target a P&L Equation, such as:
Purchase.100
Cost of Deliver...- 10
Net Sales..90
Cost of Goods. 50
Profit..40

Alternative Objectives
Classification
Proportion Correct
Diagonal

Sensitivity
Proportion correct of an observed Event
Row

Lift
Proportion correct of a predicted Event
Column

Trade Offs
Classification
none
Sensitivity
Specificity
Lift
Size of Subset (Node)

Need for a Stopping Method


Trees can grow indefinitely
Probability of an Event depends on Depth
of Tree
More Depth creates Problems to
generalise (overfitting)

Current Stopping Methods


Leaf Size
Statistical Significance of Split
Prune to best Generalization Score using
Classification Rate
Problems:
Dominant Class dominates Classification
Rare Events are a Lift Tradeoff
SAS Institute researches Stopping
Methodology

Overview of DataSplits Tool

Tool in the SAS Enterprise Miner

Icon
PROC DATASPLIT
Methods Calls (non-display AF Object
Class)

Overview of DataSplits Tool


Seamless Integration of CHAID and CART

Methodology
Nominal and Interval Targets and Inputs
Multiway Splits on all Input Types
Up to 32000 Values for a nominal Input
Allows User defined Split Criteria
General Profit and Loss Functions
Suited for Rare Events

Overview of DataSplits Tool Datasets

Overview of DataSplits Tool Variables

Overview of DataSplits Tool Options

Overview of DataSplits
Tool - Advanced Options

Overview of DataSplits Tool Decisions

Overview of DataSplits
Tool - Assessment

Overview of Tree Browser


Summary Table (upper left)
Tree-Ring Navigator (upper right)
Accessible from here: Tree Diagram +
Assessment Statistics
Assessment Table (lower left)
Assessment Graph (lower right)
blue Training Data, red Validation Data

Overview of Tree Browser

Profit Chart from


Assessment Node

Summary: Decision Trees


in SAS Enterprise Miner
Decision Trees are a powerful Tool to
segment Customer related Data
DataSplits is a very comprehensive
Implementation of Decision Trees
Decision Alternatives can be assessed
graphically and in Financial Terms
associated with each Alternative
The DataSplits Tool is one of many
integrated Tools in the SAS Enterprise
Miner.

Segmentation Using
Decision Trees
Thank you for your Attention!
Questions?

You might also like