Exploiting Context in Linear Influence Games: Improved Algorithms for Model Selection and Performance Evaluation

In the recent past, extensive experimental works have been performed to predict joint voting outcomes in Congress based on a game-theoretic model of voting behavior known as Linear Influence Games. In this thesis, we improve the model selection and evaluation procedure of these past experiments. First, we implement two methods, Nested Cross-Validation with Tuning (Nested CVT) and Bootstrap Bias Corrected Cross-Validation (BBC-CV), to perform model selection and evaluation with less bias than previous methods. While Nested CVT is a commonly used method, it requires learning a large number of models; BBC-CV is a more recent method boasting less computational cost. Using Nested CVT and BBC-CV we perform not only model selection but also model evaluation, whereas the past work was focused on model selection alone. Second, previously models were hand picked based on performance measures gathered from CVT, but both Nested CVT and BBC-CV necessitate an automated model selection procedure. We implement such a procedure and compare its selections to what we otherwise would have hand picked. Additionally, we use sponsorship and cosponsorship data to improve the method for estimating unknown polarity values of bills. Previously, only subject code data was used. This estimation must be done when making voting outcome predictions for a new bill as well as measuring validation or testing errors. We compare and contrast several new methods for estimating unknown bill polarities.