Modeling Viewer and Influencer Behavior on Streaming Platforms
Rajaram, Prashant
2021
Abstract
The video streaming industry is growing rapidly, and consumers are increasingly using ad-supported streaming services. There are important questions related to the effect of ad schedules and video elements on viewer behavior that have not been adequately studied in the marketing literature. In my dissertation, I study these topics by applying causal and/or interpretable machine learning methods on behavioral data. In the first essay, “Finding the Sweet Spot: Ad Scheduling on Streaming Media”, I design an “optimal” ad schedule that balances the interest of the viewer (watching content) with that of the streaming platform (ad exposure). This is accomplished using a three-stage approach applied on a dataset of Hulu customers. In the first stage, I develop two metrics – Bingeability and Ad Tolerance – to capture the interplay between content consumption and ad exposure in a viewing session. Bingeability represents the number of completely viewed unique episodes of a show, while Ad Tolerance represents the willingness of a viewer to watch ads and subsequent content. In the second stage, I predict the value of the metrics for the next viewing session using the machine learning method – Extreme Gradient Boosting – while controlling for the non-randomness in ad delivery to a focal viewer using “instrumental variables” based on ad delivery patterns to other viewers. Using “feature importance analyses” and “partial dependence plots” I shed light on the importance and nature of the non-linear relationship with various feature sets, going beyond a purely black-box approach. Finally, in the third stage, I implement a novel constrained optimization procedure built around the causal predictions to provide an “optimal” ad-schedule for a viewer, while ensuring the level of ad exposure does not exceed her predicted Ad Tolerance. Under the optimized schedule, I find that “win-win” schedules are possible that allow for both an increase in content consumption and ad exposure. In the second essay, “Video Influencers: Unboxing the Mystique”, I study the relationship between advertising content in YouTube influencer videos (across text, audio and images) and marketing outcomes (views, interaction rates and sentiment). This is accomplished with the help of novel interpretable deep-learning architectures that avoid making a trade-off between predictive ability and interpretability. Specifically, I achieve high predictive performance by avoiding ex-ante feature engineering and achieve better interpretability by eliminating spurious relationships confounded by factors unassociated with “attention” paid to video elements. The attention mechanism in the Text and Audio models along with gradient maps in the Image model allow identification of video elements on which attention is paid while forming an association with an outcome. Such an ex-post analysis allows me to find statistically significant relationships between video elements and marketing outcomes that are supplemented by a significant increase in attention to video elements. By eliminating spurious relationships, I generate hypotheses that are more likely to have causal effects when tested in a field setting. For example, I find that mentioning a brand in the first 30 seconds of a video is on average associated with a significant increase in attention to the brand but a significant decrease in sentiment expressed towards the video. Overall, my dissertation provides solutions and identifies strategies that can improve the welfare of viewers, platform owners, influencers and brand partners. Policy makers also stand to gain from understanding the power exerted by different stakeholders over viewer behavior.Deep Blue DOI
Subjects
Streaming Media Digital Advertising Influencer Marketing Machine Learning Deep Learning
Types
Thesis
Metadata
Show full item recordCollections
Remediation of Harmful Language
The University of Michigan Library aims to describe its collections in a way that respects the people and communities who create, use, and are represented in them. We encourage you to Contact Us anonymously if you encounter harmful or problematic language in catalog records or finding aids. More information about our policies and practices is available at Remediation of Harmful Language.
Accessibility
If you are unable to use this file in its current format, please select the Contact Us link and we can modify it to make it more accessible to you.