Song popularity prediction model based on multi-modal feature fusion and LightGBM

Abstract

Since the task of hit song prediction was proposed, many experts and technicians have done a lot of research and achieved good results, but there are still some problems such as limited song feature types, lack of feature importance, and insufficient prediction accuracy. This paper proposes a song popularity prediction model based on multi-modal feature fusion and LightGBM. In our proposed model, there is a multi-modal feature extraction structure, a LightGBM structure and a logistic regression structure. First, in order to solve the problem of limited song feature types, we fuse metadata, audio features and other relevant important features into multi-modal features. Then, in order to improve the accuracy of prediction, we introduce LightGBM algorithm to preprocess the dataset and train the model, so as to obtain the predicted value of song popularity. At the same time, we introduce a logistic regression model to research the influence of each feature on whether a song is popular from the perspective of binary classification, so that we can further study the importance of song features, and obtain the response coefficient of each feature, namely, the coefficient of response mean. Finally, we compare the prediction results of our model with the existing models, and the experiments show that the prediction results of our model have higher accuracy.