Muhammad Suleiman,
"Content-enriched Popularity Prediction for Music Tracks"
, 2024
Original Titel:
Content-enriched Popularity Prediction for Music Tracks
Sprache des Titels:
Englisch
Original Kurzfassung:
The music industry is constantly growing, driven by technological advances that are reshaping how music is produced, distributed, and enjoyed by audiences. Accurately predicting music popularity is becoming a critical ability for artists, producers, and other stakeholders in the music industry.
Previous research has mainly considered the problem of predicting music track popularity as a classification or ranking problem, often simplifying it into binary outcomes such as popular or not popular classes.
In this thesis, we take a different perspective and study the prediction of a music track's popularity, and define it as the daily percentage of listening events of a track using time series forecasting models.
Additionally, the study aims to enrich these models with various types of content features (visual, audio, textual) derived from the tracks themselves, exploring the impact of these features on the prediction process.
Two primary models are employed: DeepAR, a probabilistic Recurrent Neural Network (RNN)-based forecasting model, and Temporal Convolutional Networks (TCN), a variant of Convolutional Neural Networks (CNNs) designed to tackle sequence data. These models are initially trained using historical listening data and subsequently augmented with content features. Furthermore, a baseline Vector Autoregressive Integrated Moving Average (VARIMA) model was also utilized for comparison purposes.
Our evaluations show that while employing content features can improve the performance and predictive power of the DeepAR model, these improvements are feature and step-specific, indicating the need for careful feature engineering and selection. In contrast, the TCN model excels in scenarios with longer forecasting horizons but shows mixed results with the integration of content features. However, both models outperformed our baseline VARIMA model.