Understanding the dynamics of viewer engagement: A multi-study analysis of video features using machine learning approaches

Abstract

This thesis analyses the dynamics of viewer engagement in YouTube videos through a multi-study approach. The thesis focuses on the methodological aspect of feature extraction using python libraries and machine learning models. The features are analysed against viewer engagement data, with focus on attempting to uncover the factors that influence viewer engagement. A particular interest in the context of this dataset was to analyse the factors that make native in-content advertisements successful. The findings indicate that continuous features such as hue, luminosity and tempo have significant effects on viewer retention. Object detection analysis on the dataset also suggested multiple objects that had significant effect on viewer engagement. The data was classified into commercial and non-commercial sections manually by human-labelling. A subsequent study was conducted with a similar labelling task using a large language model where its efficacy was compared to human-labelling. This thesis provides valuable learnings for content creators with suggestions on future research directions on video content analysis and viewer engagement optimization.