Opera_vocal_classification

Muilti-classification on opera choral. The classifier should tell a opera has which vocal per second. We developed two classifiers to tell three kinds of vocal: male/ female/ choral vocal. This task has to tackle a challenge of unbalanced data.

What we do in this project:

• Data preparation: cleaned the incorrect annotations, train-test split according to the opera sequence.
• Dimention reduction.
• Adjusted the classifier for the unbalanced data problem: cost-sensitive classifiers and voting.
• Evaluation on accuarcy.
• Evaluation on profit.
• Smoothing(Not finishes in this project).

Model Performance before considering the unbalanced classes distribution

Accuarcy Estimation

Profit Estimation on Tesing Data

If it's a correct prediction, get 1 profit, otherwise get a -1 point for the profit. The goal is to earn the best profit for each classifier.

Result of Voting Classifier

The classifier compostion for voting. Final profit of the best classifier from the result of voting.