Data Mining and Milwaukee: Mining Community Health Risk Factors

1
Data Mining and MilwaukeeMining Community
Health Risk Factors

• Team Members
• Srinivas Bodapati
• Priya Hastagiri
• Dale Steber
• Madhuri Battu

2
Goal of the Project

• Create a smaller survey from the existing
  comprehensive survey

3
Data Mining Tasks

• Classification
• Build a good classifier which predicts the
  overall health status of the individuals
• Feature Selection
• Identify the meaningful attributes that impact
  the overall health status

4
Data Transformation

• 1200 instances
• 275 attributes
• Cleaned and consolidated attributes to 94
• Eliminated nulls
• Consolidated attributes that separated out based
  on gender
• Eliminated attributes that are irrelevant (Zip
  code, state, city)
• Converted the SPSS file into Excel format for
  clean up
• Cleaned data was loaded into an Access database
• Helps for easy loading in WEKA
• Modified the WEKAs jar file to add the new
  database connectivity in the jdbc connection file

5
Methods

• Methods
• Split data set into 2/3 and 1/3 sets (800
  instances training set and 400 test set)
• Classification
• ZeroR model
• OneR Model
• Decision tree( j48 Algorithm)
• Naïve Bayesian

6
Baseline Models

• ZeroR
• Accuracy 34.125
• OneR
• Accuracy 44.75

7
Model Interpretation(Before Feature Selection)
8
Attribute Selection

• Algorithms
• Information Gain
• Relief-F
• Principal Component Analysis

9
Information Gain

• Top attributes selected
• Income
• Health Care Coverage
• Education
• Depression and Stress
• Marital Status
• Exercise
• Smoking
• Alcohol Usage
• Pneumonia
• Blood Pressure and Cholesterol

10
Relief-F

• Top attributes selected
• Income
• Health Care Coverage
• Education
• Depression and Stress
• Marital Status
• Exercise
• Smoking (support of banning)
• Colonoscopy Exam
• Sunscreen
• Fast Food

11
Model Interpretation(After Feature Selection)
12
Principal Components

• Previously the analysis would not stop
• The reason WEKA did not stop was due to out of
  memory errors
• Modified the runweka.bat file to instruct the
  virtual machine to use more system memory by
  adding XmemX 1000M
• WEKA completes, but we have not had time to
  interpret the results.

13
Where do we go from here?

• Complete Principal Component Analysis
• Reevaluate representation of attributes
• Eliminate attribute with large amounts of missing
  values
• Standardize the data set
• Eliminate instances that have a class of Not
  Sure

14
Split the Data Set

• Confusion matrix
• Split data set on goal attribute
• Good, Very Good
• Poor, Fair, and Excellent

15
Confusion Matrix
16
Steps for split

• Excellent, Fair, Poor
• Repeat steps
• Very good, Good
• Apply support vector machines

17
Thank You

• Questions?