Transcript Journal of Service Research

Neural
Networks
Automatic Model
Building (Machine
Learning)
Artificial Intelligence
7-2
High-Growth Product
• Used for classifying data
– target customers
– bank loan approval
– hiring
– stock purchase
– trading electricity
– DATA MINING
• Used for prediction
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-3
Description
• Use network of connected nodes (in
layers)
• Network connects input, output
(categorical)
– inputs like independent variable values in
regression
– outputs: {buy, don’t} {paid, didn’t}
{red, green, blue, purple}
{character recognition - alphabetic characters}
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-4
Network
Input
Layer
Hidden
Layers
Output
Layer
Good
Bad
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-5
Operation
• Randomly generate weights on model
– based on brain neurons
• input electrical charge transformed by neuron
• passed on to another neuron
– weight input values, pass on to next layer
– predict which of the categorical output is true
• Measure fit
– fine tune around best fit
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-6
Operation
• Useful for PATTERN RECOGNITION
• Can sometimes substitute for
REGRESSION
– works better than regression if relationships
nonlinear
– MAJOR RELATIVE ADVANTAGE OF NEURAL
NETWORKS:
YOU DON’T HAVE TO UNDERSTAND THE MODEL
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-7
Neural Network Testing
• Usually train on part of available data
– package tries weights until it successfully categorizes
a selected proportion of the training data
• When trained, test model on part of data
– if given proportion successfully categorized, quits
– if not, works some more to get better fit
• The “model” is internal to the package
• Model can be applied to new data
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-8
Business Application
• Best in classifying data
mortgage underwriting
bond rating
commodity trading
asset allocation
fraud prevention
• Predicting interest rate, inventory
firm failure
bank failure
takeover vulnerability
stock price
corporate merger profitability
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-9
Neural Network Process
1. Collect data
2. Separate into training, test sets
3. Transform data to appropriate units
•
Categorical works better, but not necessary
4. Select, train, & test the network
•
•
•
Can set number of hidden layers
Can set number of nodes per layer
A number of algorithmic options
5. Apply (need to use system on which built)
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-10
Marketing Applications
• Direct marketing
– database of prospective customers
• age, sex, income, occupation, education, location
• predict positive response to mail solicitations
• THIS IS HOW DATA MINING CAN BE
USED IN MICROMARKETING
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-11
Neural Nets to Predict Bankruptcy
Wilson & Sharda (1994)
Monitor firm financial performance
Useful to identify internal problems, investment evaluation, auditing
Predict bankruptcy - multivariate discriminant analysis of financial ratios
(develop formula of weights over independent variables)
Neural network - inputs were 5 financial ratios - data from Moody’s
Industrial Manuals (129 firms, 1975-1982; 65 went bankrupt)
Tested against discriminant analysis
Neural network significantly better
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-12
Ranking Neural Network
Wilson (1994)
Decision problem - ranking
candidates for position, computer systems, etc.
INPUT - manager’s ranking of alternatives
Real decision - hire 2 sales people from 15
applicants
Each applicant scored by manager
Neural network took scores, rank ordered
best fit to manager of alternatives compared (AHP)
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-13
CASE: Support CRM
Drew et al. (2001), Journal of Service Research
• Identify customers to target
• Customer hazard function:
– Likelihood of leaving to a competitor
(CHURN)
• Gain in Lifetime Value (GLTV)
– NPV: weight EV by prob{staying}
– GLTV: quantified potential financial effects of
company actions to retain customers
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-14
Models:
• Proportional Hazards Regression
• Neural Networks
– Estimate hazard functions
• Baseline Regression Models
– Models for longitudinal analysis
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-15
Data: Data Warehouse of Cellular
Telephone Division
• Billing
– Previous balance, access charges, minutes used, toll charges,
roaming charges, optional features
• Usage
– Number of calls, minutes by local, toll, peak, off-peak
• Subscription
– Months in service, rate plan, contract type, date, duration
• Churn
– Binary flag
• Demographics
– Age, profitability to firm (current & future)
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-16
Model Use
• Sample of 21,500 subscribers, April 1998
• Modeled tenure for 1 to 36 months
• Trained on 15,000 of these samples
– Remainder used for testing
• Neural network models worked better than
traditional statistics
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-17
Systems
A great many products
• general NN products
$59 to $2,000
@Brain BrainMaker
Discover-It
• components
DATA MINING along with megadatabases
other products
• library
callable
• specialty products
construction bidding, stock trading, electricity trading
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved
7-18
Potential Value
• THEY BUILD THEMSELVES
– humans pick the data, variables, set test limits
• CAN DEAL WITH FAST-MOVING
SITUATIONS
– stock market
• CAN DEAL WITH MASSIVE DATA
– data mining
• Problem - speed unpredictable
McGraw-Hill/Irwin
©2007 The McGraw-Hill Companies, Inc. All rights reserved