Data Mining: Bob Stine Dept of Statistics, Wharton School University of Pennsylvania

Data Mining Introduction
Bob Stine Dept of Statistics, Wharton School University of Pennsylvania
Wharton
Department of Statistics
An insult? Predictive modeling
What is data mining?

Large, wide data sets, often unstructured Automatic, complex models
Networks, trees, ensembles black boxes Leverage modern computing, recent results in learning theory
Emphasize prediction rather than explanation

Association and prediction rather than cause and effect
Testable claims
Wharton
Science requires making claims that are testable Claims of prediction provide such a test
2
Data mining poorly matched to soc science? Response

Honest
A better match to most of us do in common practice
Data Mining in Soc Sci
Empiricism run wild, lack of theory or hypotheses Post hoc inference Need to leverage technology
Tukey Cost of theory relative to cost of computing
Diagnostic
Have I missing something?
Deep connections
Multidimensional scaling, Likelihood, Modern regr
Wharton
Week 1 Week 2 Syllabus

Wharton
Plan
Data mining with regression, logistic regression Illustrate key ideas in familiar context Alternative methods Trees, networks, ensemble methods
Boosting and bagging
Hands-on: Lab sessions on Thursdays Annotated bibliography July 4

4
Cannot learn stat without doing statistics! Modern computing provides Packages
JMP from SAS
Will be front-end to SAS Enterprise Miner Available on Newberry systems
Software
New ways to look at old things, like regression New approaches to data analysis
R Others: Stata, SPSS, Weka,

Wharton
Time series analysis
My Background
Effects of modeling on forecast accuracy Bootstrap resampling Predictive models in credit, health Alternative regression methods Combining traditional data and text Political science and voting behavior
Model selection in general Recent
Long time friend of Summer Program

Wharton
t-shirts
6
What question do you want to answer?

And do you have the right data to do it?
Research Questions
Question to guide analysis Ideal data?
Questions from science, business Social science questions

Wharton
Whos most at risk of a disease? Whats going to happen to nancial markets? Are any of these people dishonest? Will this person vote if I get them to register? Whom will this registered voter choose? Whom would those who didnt vote choose?
7
of survey Background Two waves, every two years Questions Categorical responses
2008 ANES Survey
makes these interesting What Get out the vote, phone banks
Did you vote? For whom? Numerical responses How much do you like this candidate
Wharton
ANES ideal data Is the Missing data, self-reported, interviewer effects...
Role of participation in election Would those who didnt vote change things?
90/10 rule
8
Spirit of EDA, exploratory data analysis

Know your data Know your tools
Data Browsing

ANES data table

Variable creation
Wharton
Probably several tools, using each for what it does best
Load directly from SPSS sav le 25383-0001-Data.sav Almost square: 2,323 cases x 1963 variables Virtually all categorical role of scaling Feeling thermometers, moderators (N5)
No algorithm as good (yet) as the modeler who knows how to build predictive features
9
Marginal distributions Interesting variables
Browsing ANES
Interactive graphics: Plot linking and brushing Participation, political interest (A1-A10)
prevalence of missing data. Problem for categorical data?
Feeling thermometer (B1 group)

numbers or categories? Missing a problem?
JMP treatment of numerical/ categorical
Spending bundle and scaling (P1 group)

Likert scales, ordinal-interval-ratio measurement
Intention to vote (A6, Q1 in rst wave)

Repeats prior question, reliability of data
Choice in election (C6,Q14 in second wave)

Wharton
Importance of sampling weights (65.5% in sample, 52.9% in election)
10
Bivariate relationships
Browsing ANES
Special scatterplot if mix types
Contingency tables, scatterplots Asymmetry of roles: explanatory vs response
Consistency of responses
: scatterplot
Choice and opinion of war in Iraq Choice and rating of candidate
FT rating of Dem candidate pre/post election
B1/D1
War and voter choice: table, mosaic plot
Q14/A14f
Feelings and voter choice: logistic regression

Wharton
Q14/D2xx
11
Models in statistics Example: SRM Inference

Wharton
Models
All models are wrong, some are useful Box
What is a model? Why are they needed? E(Y|X) = 0 + 1X + independence, equal var normal Statistical signicance Condence interval, test, p-value
Interpretation?
12
Residual diagnostics Calibration
Model Diagnostics
Is the model correct on average Check by smoothing Y on X or Y on
Interactive tool for spline in JMP
Wharton
13
Does one explanatory variable a complete

description of the response?
Media Emotional interest in outcome Attitude to Iraq war, economy,
Multiple Regression
What other factors affect association between pre-election rating and post rating?
How do these factors contribute to model

Additive as another explanatory variable Affecting other factors (interaction)
How should we decide which?

Wharton
Trial and error by adding to multiple regression? Use of t-statistics and p-values to decide
14
Multiple Regression Model

Underlying model has assumptions Key assumption is the larger equation Xs are known
E(Y|X) = 0 + 1 X1 + 2X2 + + kXk and additive Same assumptions for the unexplained variation
Grow to a multiple regression model
Judging the other variables
Which do we keep, which do we exclude? How many variables did you try? What made you try those? What about other correlated variables?
Use of t-statistics, F-statistics in this setting

Wharton
15
Grow simple regr into a multiple regression

model that includes interactions
Add Happy/Care, care who wins Interaction What does all of this tell you?
Possible Model
Wharton
16
Prole of Model Alternative way to look at a model

Visual presentation of effects vs tabular What does the interaction do? Bars indicate condence
Error bars indicate condence
Wharton
17
Surface prole
Looking at Fit
How would it look were there no interaction?
Wharton
18
Role of data mining in social science research

Diagnostic Better way to do what we are doing Right on average
Take-Aways
Calibration of a model Classical statistics, inference Interactive graphical visualization

Wharton
Exploring data (plot linking, brushing) Exploring models (proling, surfaces)

19
Skim syllabus, bibliography Think about what you did for models Peek at the codebook for ANES
Z drive on Newberry computers Come with questions
Assignment
Think about your own modeling and data
Wharton
20
Picking the variables for a regression What about those missing values?
Next Time
Wharton
21

Data Mining: Bob Stine Dept of Statistics, Wharton School University of Pennsylvania

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Data Mining: Bob Stine Dept of Statistics, Wharton School University of Pennsylvania

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Data Mining: Bob Stine Dept of Statistics, Wharton School University of Pennsylvania

Uploaded by

Copyright:

Available Formats

Data Mining Introduction

Bob Stine Dept of Statistics, Wharton School University of Pennsylvania

An insult? Predictive modeling

What is data mining?

Emphasize prediction rather than explanation

Data mining poorly matched to soc science? Response

Data Mining in Soc Sci

Week 1 Week 2 Syllabus

Hands-on: Lab sessions on Thursdays Annotated bibliography July 4

R Others: Stata, SPSS, Weka,

Time series analysis

Model selection in general Recent

Long time friend of Summer Program

What question do you want to answer?

Questions from science, business Social science questions

2008 ANES Survey

ANES ideal data Is the Missing data, self-reported, interviewer effects...

Spirit of EDA, exploratory data analysis

ANES data table

Probably several tools, using each for what it does best

Marginal distributions Interesting variables

Feeling thermometer (B1 group)

JMP treatment of numerical/ categorical

Spending bundle and scaling (P1 group)

Intention to vote (A6, Q1 in rst wave)

Choice in election (C6,Q14 in second wave)

Importance of sampling weights (65.5% in sample, 52.9% in election)

Contingency tables, scatterplots Asymmetry of roles: explanatory vs response

FT rating of Dem candidate pre/post election

War and voter choice: table, mosaic plot

Feelings and voter choice: logistic regression

Models in statistics Example: SRM Inference

All models are wrong, some are useful Box

Residual diagnostics Calibration

Is the model correct on average Check by smoothing Y on X or Y on

Interactive tool for spline in JMP

Does one explanatory variable a complete

How do these factors contribute to model

How should we decide which?

Multiple Regression Model

Grow to a multiple regression model

Judging the other variables

Use of t-statistics, F-statistics in this setting

Grow simple regr into a multiple regression

Prole of Model Alternative way to look at a model

Error bars indicate condence

How would it look were there no interaction?

Role of data mining in social science research

Calibration of a model Classical statistics, inference Interactive graphical visualization

Exploring data (plot linking, brushing) Exploring models (proling, surfaces)

Think about your own modeling and data

You might also like