Article

Quantifying the Dominant Role of Channel Audience in Predicting YouTube Success with Machine Learning 

---

Abstract

This study investigates the determinants of YouTube video virality by quantifying the relative influence of audience capital and metadata optimization using machine-learning techniques. Employing a quantitative, cross-sectional design, metadata from the top 100 YouTube music videos of 2025 was analyzed to predict video popularity, measured by view count. After data preprocessing and feature engineering, two supervised regression models, Multiple Linear Regression and Random Forest Regressor were trained and evaluated. The linear model performed poorly (R² = –0.0680; RMSE = 1.0970), indicating that linear assumptions failed to capture the complex relationships among features. In contrast, the Random Forest model achieved strong predictive performance (R² = 0.7157; RMSE = 0.5660; OOB = 0.7051) and provided interpretable feature-importance metrics. Results revealed that channel_follower_count overwhelmingly dominated all other predictors, accounting for over 80% of the total importance score, while metadata-related variables such as description_length, tag_count, and categories contributed marginally. These findings empirically support the thesis that a creator’s accumulated audience capital is the primary determinant of virality, reflecting the self-reinforcing dynamics of the digital attention economy. The study highlights the limitations of content-level optimization strategies and underscores the structural advantages enjoyed by established creators in algorithmic ecosystems. Implications extend to platform governance, creator strategy, and digital inequality, emphasizing the need for more transparent and equitable recommendation systems.

Keywords: YouTube virality, machine learning, platform capital, algorithmic visibility, audience engagement

How to Cite:

Ramadani, N. C. & Fahreza, S., (2025) “Quantifying the Dominant Role of Channel Audience in Predicting YouTube Success with Machine Learning ”, Journal of Digital Society 1(4), 258-271. doi: https://doi.org/10.63913/jds.v1i4.41