Predicting Political Donations Using Twitter Hashtags and Character N-Grams

We describe a novel approach for predicting politicaldonations and performing psychographic segmentation basedon social data linked to election donation records. The role ofmicroblogs in enterprise informatics, specifically in relation tocustomer relationship systems is highlighted. Algorithms trainedon social data can be used to interpret and detect prospects' psychographicinformation. Contrasted with past approaches whichfocused exclusively on a single source of social data, the methodbeing presented allows us to use an objective gold standard bylinking Twitter and election records. Two experiments were conductedusing data collected from 438 Twitter users, half of whichare linked with donation event records collected from the UnitedStates Federal Election Commission. Probabilistic, entropy andkernel approaches were tested for predictive accuracy, while theCNG technique is explored as an alternative. The CNG algorithmwas found to predict political affiliation 17 percentage pointsabove the majority classifier, exceeding benchmarks suggestedby the literature. A NaïveBayes word n-gram approach wasfound to outperform CNG at predicting donations by predictingpolitical donations. Insufficient performance and poor reliabilityof standard word n-gram techniques in opinion detection revealskepticism about past work on political affiliation analysis fromsocial data alone. This suggests that prospecting systems maybenefit from constructing algorithms using data linked to external sources.