Slide 1AbstractVoting RepresentationsExperimental ProceduresExperimental ParametersExperiment 1 – Network StructureExperiment 2 - CoefficientsClassification ResultsConcluding RemarksPrediction of Voting Patterns Based on Census and Demographic DataAnalysis Performed by: Mike HeECE 539, Fall 2005AbstractAbstractPrediction of Voting Patterns in 2004 Prediction of Voting Patterns in 2004 Presidential ElectionPresidential ElectionMulti-Layer Perceptron, Back-PropagationMulti-Layer Perceptron, Back-PropagationBased on Demographic DataBased on Demographic DataPopulation SizePopulation SizeGender CompositionGender CompositionRacial CompositionRacial CompositionAge CompositionAge CompositionVoting RepresentationsVoting RepresentationsArea-Based Winner- Takes-All Map•Strict Red/Blue binary color coding•Can misrepresent actual popular opinionPopulation-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 DemocraticLinearly Shaded Vote-Percentage Map•Colors shaded according to vote percentages•Accurately portrays closeness of most races and political homogeneity throughout countryExperimental ProceduresExperimental ProceduresData Pre-ProcessingData Pre-ProcessingNetwork Structure DeterminationNetwork Structure Determination# of Hidden Layers, Neurons in Layers# of Hidden Layers, Neurons in LayersCoefficients DeterminationCoefficients DeterminationTraining, Training Error TestingTraining, Training Error TestingError from vote percentages, calling for Error from vote percentages, calling for candidatecandidateTesting on Testing Data SetTesting on Testing Data SetExperimental ParametersExperimental Parameters14 Features, 3 Outputs14 Features, 3 OutputsHyperbolic Tangent Activation Function for Hyperbolic Tangent Activation Function for Hidden LayersHidden LayersSigmoid Activation Function for Output Sigmoid Activation Function for Output LayerLayerLearning coefficient Learning coefficient αα=0.2=0.2Momentum coefficient Momentum coefficient μμ=0.5=0.5Experiment 1 – Network StructureExperiment 1 – Network StructureMany different structures tested according Many different structures tested according to total square errorto total square errorBest performers isolated for further testingBest performers isolated for further testingComparison of error across multiple trials Comparison of error across multiple trials between tested structuresbetween tested structuresWinner: 15 neurons in hidden layer, 4 Winner: 15 neurons in hidden layer, 4 hidden layershidden layersExperiment 2 - CoefficientsExperiment 2 - CoefficientsTo determine optimum To determine optimum αα and and μμDifferent sets of coefficients tested based Different sets of coefficients tested based on total square error as well as maximum on total square error as well as maximum square errorsquare errorChosen configuration:Chosen configuration:αα = 0.2, and = 0.2, and μμ = 0.5 = 0.5Classification ResultsClassification ResultsApplication of MLP to attempt to predict Application of MLP to attempt to predict which candidate will win each countywhich candidate will win each county100 training and prediction trials100 training and prediction trialsFor Wisconsin (training data), 77% For Wisconsin (training data), 77% classification rateclassification rateFor Minnesota (testing data), 75% For Minnesota (testing data), 75% classification rateclassification rateLess than 3% standard deviation in Less than 3% standard deviation in classification rate between trialsclassification rate between trialsConcluding RemarksConcluding RemarksImpressive overall predictive powerImpressive overall predictive powerRetains predictive power for different states:Retains predictive power for different states:Wisconsin and Minnesota similar demographically, Wisconsin and Minnesota similar demographically, different politicallydifferent politicallyPredictions based only on demographics – Predictions based only on demographics – innocuous data leads to powerful resultsinnocuous data leads to powerful resultsDemonstrates effectiveness of MLP’s as well as Demonstrates effectiveness of MLP’s as well as element of truth in common generalizations of element of truth in common generalizations of demographic voting tendenciesdemographic voting
View Full Document