Application of Machine Learning on NBA Data Sets

Abstract

Machine learning is known as the most popular methodology to do prediction on large data set while NBA's data sets consists of plentiful statistics. Since predictions of various events are important, our research would investigate whether machine learning algorithms are efficient in doing prediction on certain NBA data sets and tasks. We are focus on mainly three supervised tasks, namely: All-Star Prediction, Playoff Prediction and Hot Streak Fallacy. For Playoff Prediction, we predict the team performance by doing machine learning on two data sets consisting of distinct well-selected features and compare the result to show which data set are more suitable for the machine learning to work. The results show that advanced statistics outperform the elementary ones. For Hot Streak Fallacy, we build the model based on multiple-linear regression to address the question: is hot streak a fallacy? It turns out that there is a lack of evidence to support 'Hot Streak Phenomenon'. For the NBA Trend, we try to view how the games involve for the past decade, and analyze the correlation of playoff tickets and other data.