Pairwise Learning to Rank for Hit Song Prediction

Abstract

Predicting the popularity of songs in advance is of great interest to the music industry, with possible applications including assessing the potential of a new song, automated songwriting assistants, or song recommender systems. Traditional approaches for solving this use pointwise models focused on single songs, either using classification to categorize songs into classes like hit and non-hit, or regression to predict popularity metrics like play count. We propose to draw inspiration from research on learning to rank and instead use a pairwise model. Our model takes a pair of songs A and B and predicts whether song A is more popular than song B. Based on this problem formulation, we propose a neural network model that is trained in a pairwise fashion, as well as two data augmentation strategies for improving its performance. We also compare our model to one trained in a traditional pointwise way. Our results show that the pairwise model using our proposed augmentation strategies outperforms the pointwise model.