ΣΥΣΤΗΜΑ ΣΥΛΛΟΓΗΣ ΚΑΙ ΕΠΑΝΑΔΙΑΘΕΣΗΣ ΔΕΔΟΜΕΝΩΝ ΔΗΜΟΤΙΚΟΤΗΤΑΣ ΜΟΥΣΙΚΩΝ ΚΟΜΜΑΤΙΩΝ ΚΑΙ ΠΡΟΒΛΕΨΗ ΤΟΥΣ ΜΕΣΩ ΑΚΟΥΣΤΙΚΩΝ ΧΑΡΑΚΤΗΡΙΣΤΙΚΩΝ
Date Issued
July 16, 2021
Type
Πτυχιακή Εργασία
Abstract
In this work, we attempt to solve the ”Hit Song Science” problem which aims to the prediction of commercially successful songs before their distribution using their audio features and characteristics. Predicting hit songs would have a huge impact on the development of the music industry. Artists would be able to focus their efforts on more promising and maybe even better songs and therefore increase their profits. Also further analyzing data from popularity sources would lead to a better understanding of the history and trends of music. Several works in the past have tried to address the issue. Such works exist since 2005. Our first part of the work was based on a previous work that was published in 2016. In the first part of this work, we constructed a database consisting of 2844 Music charts, 22600 hit songs, and 8329 artists, dating back to 1985 till today. Three popularity sources were used to collect the data namely billboard, Spotify, and shazam. Also, an automated data collection system was constructed. Attention was given to its error-proofing and speed in an attempt to make a fully automated weekly update system for our database. Later a website was made to present our work. The database was used on our website to provide a user-friendly way to explore the data through various types of diagrams. For example, a user can explore the presence an artist had on music charts or explore all of his songs that made it to the charts. In the second part of this work, we constructed a dataset in an attempt to solve the Hit Song Science problem. First using the database we mentioned earlier we constructed a dataset with approximately 17430 hit songs. Later we constructed a second dataset with approximately 17315 nonhit songs using a third party dataset. We then combined the two datasets and extracted the audio features for every track using the Spotify API. The column ”On chart” was also added to indicate if the song is a hit or not. The final dataset contained approximately 34740 tracks and 17 features for each of the songs. To predict whether a song will be a hit or not we tested four models on our dataset. Before testing the models we analyzed the dataset to understand our data and check for missing or invalid values that may existed. The analysis of the dataset showed a strong influence of the ”instrumentalness” and ”valence” features on the ”On chart” indicator. This led us to the assumption that the amount of vocals in the song and the musical positiveness conveyed by the song may have a serious impact in defining a song as hit or non-hit. Then we normalized our data using the min-max normalization technique to make it more suitable for testing. The four models we used were:
Support Vector Machine
Neural Network
Random Forest
k-NN
The best model was Neural Network, which was able to predict hit songs with 78.71% accuracy. The four models had the following accuracy results.
k-NN SVM Neural Networks Random forest
75.68 75.43 78.71 76.03
It should also be considered that according to the confusion matrix we constructed later, Random Forest algorithm had fewer False Positive classifications than any other algorithm. Finally, we believe that in the future more chart sources should be considered. Also, further attention should be given to ”instrumentalness” and ”valence” audio features which seemed to have a strong influence on the class.
Subjects