M.S. Applied Data Science - Capstone Chronicles 2025

11

structural and temporal metadata relate to audience interaction. The findings underscore the complexity of digital engagement and suggest that predicting video performance requires more than just surface-level features. Although the metadata used in this study did not yield strong model performance, it still proved useful for uncovering trends and generating insights that may guide future content decisions. This work provides a foundation for future research that can build on these results by integrating richer data sources and more advanced techniques. 6.2 Recommend Next Steps/Future Studies Although this study explores the potential use of metadata and text-based features to predict YouTube video popularity, the limitations of the models’ performance suggest a few opportunities for future research, improvement, and expansion. First, future studies should incorporate richer and dynamic features that have demonstrated relevance to the content performance. For example, incorporating channel-based features such as total videos uploaded, subscriber count, or total view count can provide stronger signals toward the performance of a future content upload. Similarly, incorporating video-specific attributes such as video length, tags, and even thumbnail-specific information can provide more insights into the video content itself. In addition to focusing on feature engineering, using sentiment analysis can provide further insights into the text and correlation to higher engagement. The use of certain words or phrases can capture the curiosity of people more than other words or phrases. Therefore, knowing how linguists play a role in content engagement can also provide helpful insights.

From a modeling perspective, the use of deep learning models can allow better generalization on the text-based features, especially with models such as transformers. In parallel with this, further expanding the dataset can improve performance, particularly when paired with more complex models. The use of deep learning can provide rich text-based embeddings and sequential user behavior data, allowing the capture of different complex relationships between text-based and metadata features. Finally, future research should explore how platform-level signals, such as recommendation placement, watch history, and algorithmic amplification, influence engagement. These dynamics may explain much of the unexplained variance observed in this study. Access to such contextual information, although more difficult to obtain, could be critical for developing more reliable forecasting tools. Together, these directions highlight the need for a multi-layered approach that combines metadata with behavioral, contextual, and content-driven features to better understand and predict video popularity on platforms like YouTube. ACKNOWLEDGMENTS Gratitude is extended to Dr. Ebrahim Tarshizi for his continuous guidance, feedback, and encouragement throughout this capstone project. Appreciation is also given to the faculty and staff of the ADS program for providing the foundational knowledge and tools necessary to carry out this research. The authors would like to acknowledge the use of ChatGPT (OpenAI, 2025) for support with grammar checking, technical writing assistance, and language refinement during the preparation of this report.

123