The decision process of the electoral college can be seen as a machine learning problem where the input consists of a vector of length 51 describing the election outcome in 50 states and the District of Columbia and where the final result is given as an output scalar equal to either +1 or -1. Support vector machines are applied in this context to a set of simulated and historic US presidential elections. The decision surface can best be modelled by a separating hyperplane where the weights of the hyperplane represent a scaled level of importance for each of the 50 states and the District of Columbia. For 100 simulations of 100 elections, in which Democrats and Republicans had a 50% probability of winning each state, support vector machines appear to retrieve a large-state bias. For a historic dataset covering the last 22 U.S. presidential elections no large-state bias could be detected.
[1]
Lawrence D. Longley,et al.
The Biases of the Electoral College in the 1990s
,
1992,
Polity.
[2]
Stuart Elaine Macdonald,et al.
The Power of the States in U.S. Presidential Elections
,
1986,
American Political Science Review.
[3]
H. Margolis.
The Banzhaf Fallacy
,
1983
.
[4]
Carleton W. Sterling.
The Electoral College Biases Revealed: the Conventional Wisdom and Game Theory Models Notwithstanding
,
1978
.
[5]
J. Kallenbach.
Our Electoral College Gerrymander
,
1960
.
[6]
S. Gunn.
Support Vector Machines for Classification and Regression
,
1998
.
[7]
Banzhaf,et al.
One Man, 3.312 Votes: A Mathematical Analysis of the Electoral College
,
1968
.