Model

​

Our model makes rankings by predicting the numerical “winner score” (in our training data, a song’s score was 1 if the song won Record of the Year and was 0.2 if the song was nominated for the award) for each song in a given set using a multilayer perceptron model with two hidden layers. To reflect the greater predictive power of recent results, our training data was also weighted using a logistic function in the difference between each song’s year of release and the last year in our training set. Songs with higher predicted scores are predicted as more likely to win Record of the Year.

 

To test the performance of our model, we trained on songs from 1958-2010 and generated predictions for the scores of songs from 2011-2015. Using these scores, we created rankings of songs by their year of release and examined the ranked positions of the Record of the Year winners and nominees in each year to evaluate our model. The graph below displays our results. Our model correctly predicted the Record of the Year winner in 2011, 2012, and 2014, while ranking 2013’s winner 5th and 2015’s winner 21st.