Unbox the Black-Box: Predict and Interpret YouTube Viewership Using Deep Learning

ABSTRACT

As video-sharing sites emerge as a critical part of the social media landscape, video viewership prediction becomes essential for content creators and businesses to optimize influence and marketing outreach with minimum budgets. Although deep learning champions viewership prediction, it lacks interpretability, which is required by regulators and is fundamental to the prioritization of the video production process and promoting trust in algorithms. Existing interpretable predictive models face the challenges of imprecise interpretation and negligence of unstructured data. Following the design-science paradigm, we propose a novel Precise Wide-and-Deep Learning (PrecWD) to accurately predict viewership with unstructured video data and well-established features while precisely interpreting feature effects. PrecWD’s prediction outperforms benchmarks in two case studies and achieves superior interpretability in two user studies. We contribute to IS knowledge base by enabling precise interpretability in video-based predictive analytics and contribute nascent design theory with generalizable model design principles. Our system is deployable to improve video-based social media presence.

KEYWORDS:

• Design science
• deep learning
• video prediction
• analytics interpretability
• unstructured data

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplemental data

Supplemental data for this article can be accessed online at https://doi.org/10.1080/07421222.2023.2196780

Notes

¹ In May 2020, a video called “Plandemic” featured a prominent anti-vaxxer falsely claiming that billionaires were helping to spread the virus to increase use of vaccines. By the time YouTube removed the video, it had already hit 7.1 million views [63]. Other examples are in online supplementary appendix 1.

² This unit effect is consistent with the interpretation format of linear regression. Although the prediction capability of linear regression is weak, it offers an easily understandable and largely accepted interpretation mechanism. The weight ${\beta }_i$ of a variable is usually interpreted as when $X_i$ increases one unit, $y$ will increase $i$. This unit effect format has been commonly adopted in many interpretable machine learning studies for various applications [24]. Readability is the Flesch Reading Ease, formulated as: $206.835-1.015\left(\frac{totalwords}{totalsentences}\right)-84.6\left(\frac{totalsyllables}{totalwords}\right)$, which is the most popular and the most widely tested and used readability measurement by marketers, research communicators, and policy writers, among many others. Increasing readability means using fewer words in a sentence and using words with fewer syllables.

³ After the survey, we disclosed how their model performed in relative to the other four models. We compensated them with different-valued office supplies in the end, according to the model performance ranking.

Additional information

Funding

This research was carried out with the support of the “University of Delaware General University Research” fund. Yidong Chai is supported by National Natural Science Foundation of China (72293581, 91846201, 72293580, 72188101).

Notes on contributors

Jiaheng Xie

Jiaheng Xie ([email protected]) is an Assistant Professor in the Department of Accounting & MIS at the University of Delaware’s Alfred Lerner College of Business and Economics. His research interests are interpretable deep learning, health risk analytics, and business analytics. His prior works have been published in premier journals, including MIS Quarterly and Journal of Management Information Systems.

Yidong Chai

Yidong Chai ([email protected]; corresponding author) received his PhD at Tsinghua University, China. He is a researcher in the School of Management of Hefei University of Technology, Philosophy and Social Sciences Laboratory of Data Science and Smart Society Governance of Ministry of Education, and Key Laboratory of Philosophy and Social Sciences for Cyberspace Behaviour and Management, in China. Dr. Chai’s research interests include machine learning, cybersecurity, business intelligence, and health informatics.

Xiao Liu

Xiao Liu ([email protected]) is an Assistant Professor in the Department of Information Systems at Arizona State University. She received her PhD in Management Information Systems from the Eller College of Management at the University of Arizona. Dr. Liu’s research interests include data science and predictive analytics in healthcare, education, and fintech. Her work has appeared in several academic journals and peer-reviewed conferences, such as MIS Quarterly, Journal of Management Information Systems, Journal of Medical Internet Research, Journal of the American Medical Informatics Association, and the Proceedings of International Conference in Information Systems, among others.