A machine learning model that predicts 'hit songs' with 87% accuracy is born
By
Two students from the University of San Francisco have announced a study on ' Predicting songs to be hit by machine learning' using the various data sets 'Spotify Web API' published by Spotify.
[1908.08609] Song Hit Prediction: Predicting Billboard Hits Using Spotify Data
https://arxiv.org/abs/1908.08609
Using Spotify data to predict what songs will be hits
https://techxplore.com/news/2019-09-spotify-songs.html
Kai Middlebrook and Kian Sheik are studying machine learning at the University of San Francisco. The two loved music as much as they listened to music all day, and one day they talked about music, and the topic that “hit songs sound similar” was raised. In order to verify this idea, the two perform the tempo, key, tune, valence, energy, and acousticity of the song that Spotify has published in the Spotify Web API. Analyzing ease of performance (danceability) and loudness by machine learning, we started research on ' searching for conditions of hit songs ' and ' predicting which song hits '.
By
The two analyzed data for 1.8 million songs published on the Spotify Web API using four types of algorithms: logistic regression , neural network , support vector machine, and random forest . We created machine learning models to derive prediction results of “hit” or “no hit” from various features of the song.
When we conducted a verification experiment “Predicting whether songs in the top 100 hits on the billboard will be hit with four types of machine learning models”, the results shown in the table below were obtained. The evaluation items were accuracy , accuracy, and recall . The random forest model achieved 88.7% and 87% in accuracy and accuracy, respectively, and the logistic regression model had a recall of more than 92%. .
According to Middlebrook's explanation, the accuracy indicates “% of the expectations of“ hit ”and“ no hit ”was actually correct”, and the accuracy is “% of all hit songs hit” For music labels, etc., a model with excellent accuracy is the most practical.
This research can be said to be 'predictable whether the song will hit based on the characteristics of the song'. But on the other hand, Shiek said, “This hit song prediction machine learning model is useful for the label, but if you rely on this model too much, all songs may be similar. YouTube etc. use machine learning algorithms In order for the created movie to be displayed in the recommended video, there is a criticism that it must be a uniform work, ' Expressed concern.
By Vova Krasilnikov
Middlebrook and Sheik said they plan to investigate the impact of social media, artist awareness, labels, etc. on 'whether the song will hit'.
Related Posts:
in Science, Posted by darkhorse_log