Prediction of Voting Patterns Based on Census and Demographic Data Analysis Performed by: Mike He ECE 539, Fall 2005
Abstract Prediction of Voting Patterns in 2004 Presidential Election Multi-Layer Perceptron, Back-Propagation Based on Demographic Data Population Size Gender Composition Racial Composition Age Composition
Voting Representations Area-Based Winner- Takes-All Map Strict Red/Blue binary color coding Can misrepresent actual popular opinion Population-Based Winner- Takes-All Cartogram Counties resized to reflect actual population More accurately reflects popular opinion Illustrates high density of urban areas and tendency to vote Democratic Linearly Shaded Vote- Percentage Map Colors shaded according to vote percentages Accurately portrays closeness of most races and political homogeneity throughout country
Experimental Procedures Data Pre-Processing Network Structure Determination # of Hidden Layers, Neurons in Layers Coefficients Determination Training, Training Error Testing Error from vote percentages, calling for candidate Testing on Testing Data Set
Experimental Parameters 14 Features, 3 Outputs Hyperbolic Tangent Activation Function for Hidden Layers Sigmoid Activation Function for Output Layer Learning coefficient α=0.2 Momentum coefficient μ=0.5
Experiment 1 – Network Structure Many different structures tested according to total square error Best performers isolated for further testing Comparison of error across multiple trials between tested structures Winner: 15 neurons in hidden layer, 4 hidden layers
Experiment 2 - Coefficients To determine optimum α and μ Different sets of coefficients tested based on total square error as well as maximum square error Chosen configuration: α = 0.2, and μ = 0.5
Classification Results Application of MLP to attempt to predict which candidate will win each county 100 training and prediction trials For Wisconsin (training data), 77% classification rate For Minnesota (testing data), 75% classification rate Less than 3% standard deviation in classification rate between trials
Concluding Remarks Impressive overall predictive power Retains predictive power for different states: Wisconsin and Minnesota similar demographically, different politically Predictions based only on demographics – innocuous data leads to powerful results Demonstrates effectiveness of MLP’s as well as element of truth in common generalizations of demographic voting tendencies