Classifiers

We used Weka to experiment with different classifiers and identify which would perform best. Using 10-fold cross-validation on the entire dataset, the accuracies of each classifier are displayed in the following table. We chose the following classifiers because they can be used to map continuous input attributes to categorical classifications, and we learned about how they work from EECS 349 lecture/assignments.

10-fold cross-validation accuracy for various learners. k Nearest Neighbor has peak accuracy at k=6, which is shown in the table.

ZeroR gives a baseline accuracy to compare other classifiers. It is a helpful baseline because it assigns classifications based on the the most common classification value, essentially creating a measure of minimal performance. Also, it should be noted that ZeroR is only slightly better than randomly guessing, where there is only a 1/7 chance of guessing correctly; this is because the training data was balanced for each genre.

Random forests are an ensemble learning method that involves constructing a number of individual decision trees and averaging their results. These classifiers use bagging and random selection of features to generate highly accurate models. Bagging also reduces overfitting in these trees.