A tale of machine learning based YouTube ad view prediction: analysis and interpretation

YouTube is the second-highest-used social media platform for online marketing users and social media influencers. Monetization of user video depends on the view counts, and it closely associates with the video content. Generation of viral videos with millions of views is desirable, and thus Machine Learning (ML) has emerged as a disruptive technology for ad view prediction. However, the analysis of these ML algorithms is highly crucial for understanding the semantics of video creation and dissemination. To date, limited works have addressed this problem. In our work, we present a comparative analysis of popular ML algorithms for this problem and compare them to the standard benchmark YouTube Kaggle dataset. The dataset contains around 15,000 YouTube videos, specifically curated from different regions, and of different time duration. We consider the video length, likes, dislikes, and views as key parameters and use train the dataset on five ML models, namely, the Linear Regression, Decision Tree, Support Vector Machines, Random Forest Regressor, and Artificial Neural Network (ANN) for the analysis. For performance analysis, metrics like accuracy and mean absolute error (MAE) are considered. Our analysis forms conclusive indicators for advertisers and creators to optimize their marketing campaigns based on the model outputs, which helps them to generate high monetization from their content.