Wednesday, January 23, 2002

posted 1/23/2002 03:48:00 PM by Daniel   0 comments

I just had a brainstorm for a great political science project or thesis. There have been lots of attempts to categorize people's politics, from the simplistic "left <---> right" axis to the somewhat more sophisticated 2-dimensional "Nolan Chart" that charts economic freedom on one axis and personal freedom on the other axis. I've seen many variations of the Nolan Chart, each usually containing 5 or 10 questions per axis, which supposedly pinpoint the respondent's political coordinates from the resulting scores.

But which questions are the best ones with which to grade a person's politics? How well do they really distinguish typical political positions? What kind of predictive power do they offer?

My idea is to use neural networks to classify people's politics: Come up with a questionaire containing a fairly lengthy and diverse set of questions, covering a wide variety of topics. Then recruit people to answer the questionaire who are widely recognized as epitomizing different political categories. For instance, recruit many prominent libertarians and Libertarian Party leaders who most everyone agrees hold hard-core libertarian beliefs. Do the same for environmentalists and Green Party leaders. Do the same for died-in-the-wool liberals, arch-conservatives, Republicans, Democrats, socialists, anarchists, foreign-policy "hawks" and "doves", etc.

These initial questionaire respondents provide the "training vectors" for the neural network. The neural network learns to separate people into political categories based on the complex weightings and interrelations of the answers they give to the various questions. Post-analysis of the neural network may also suggest certain questions or combinations of questions that offer great discriminatory value, while other questions and issues turn out to be largely irrelevant or derivative. With enough testing and optimizing of questions, it should be possible to classify people according to political labels that will have some real validity.

This could ultimately lead to very practical political applications, which for now I will leave to the reader's imagination.