LearnedLeague is an online trivia competition where players compete against one opponent each day to answer six questions.
This project looked to use machine learning to predict whether or not a player would get a specific question correct based on their answers to past questions. A stocastic gradient descent algorithm with one hidden layer and L2 weight regulation. This was implemented using Pytorch only as a linear algebra library (and not for its machine learning capabilites) since this was meant as an educational refresher exercise for me.
The question chosen to predict was 'Beale fired - Obscentities Prokoke Record Calls' is the hedline of a fictional 1975 edition of the New Work Daily News that appears in what film? (the answer is, of course, Network). This question was chosen since close to half (53%) of players got the answer correctly, meaning that any testing accuracy over 53% would be better than random guessing. Ultimatley, the algorithm produced a testing accuracy of 81%, meaning that based on a player's answers to 599 other, totally unrelated questions, we can get a pretty good idea if they would have known the answer to this question about Network.