A pattern classifier based approach to Campaign planning Gerrymandering A pattern classifier based approach to Campaign planning
Goals Accurate prediction of votes in a region Considering Campaign stops Considering Advertising budget Inexpensive implementation No need for an on-hand ANN expert No need to purchase expensive software package No need to run on high-end workstation “User-friendly” interface Minimum of human interaction Results are already interpreted, and presented in text-based format
Implementation Implemented in GNU Octave Mixture of experts-type system Result is weighted average of sub-classifiers Weights are based on final Crate from training Currently Implemented MLP KNN (order 3) GMM In progress Fuzzy Classifier SOM-based classifier
Data Currently Have Want Census data from 2000 and 1990 Obtained from http://www.census.gov/ Vote results from 1850-2000 (all presidential elections) Obtained from National Records Archive Want Number of campaign stops per state (2000 and 1990) Advertising dollars per state (2000 and 1990)
Difficulties (1/2) Census Campaign budget data Different data collected Stored in different format Distributed through still different means Not enough disk quota (on CAE workstations) or hard drive (on my PC) to store all of the data Campaign budget data Hard to collect (no central repository) Differing formats Usually not complete
Difficulties (2/2) Data is hard to parse No way to script retrieval Some scripting of processing, but it still requires a lot of human interaction Census data and vote data need to be manually combined Vote data must be manually extracted from HTML tables