Improving monetization in mobile advertising using machine learning

Professor De Reyck and his team developed ad-serving algorithms that increased revenue by 23% at Vungle Inc., one of the leading mobile advertising companies.

When a customer enters an advertising network, the vast amount of data available creates opportunities for advertisers to target the customer with a granularity that was impossible before, serving a personalised advertisement predicted to have maximum impact. However, utilizing the massive amount of available data to implement the most effective advertising campaign in real time is a nontrivial problem, further complicated by the fact that ad-serving decisions often need to be made in a matter of milliseconds to ensure a smooth user experience. 

Professor Bert De Reyck and his team developed ad-serving algorithms based on machine learning models, which calculate the expected profit from displaying each ad and make allocation decisions, all in less than 50ms to enable an instantaneous response. It incorporates a user-specific, real-time procedure that strikes a balance between sending a large variety of ads and sending ads of high quality, considering the user’s level of engagement with the host application. This diversification mechanism exposes users who are highly engaged with the host application to a larger variety of ads, by considering their recent interactions with ads and signs of fatigue. For a user with lower engagement, the algorithms rotate ads at a lower speed, and gives higher selection probabilities to higher-quality ads. The popularity of mobile games changes rapidly, and classification algorithms should consider fresh data that represent those changes. By implementing a two-stage training approach, the algorithm dramatically shortens training times and therefore incorporate more frequent data updates.  

The ad-serving system was first implemented by Vungle in 2014 and is still in operation today. Professor De Reyck’s team performed several A/B (comparative) tests to compare the algorithm’s performance with that of Vungle’s legacy algorithm. The results showed an immediate 23% increase in revenue, equivalent to an increase in revenue of more than $1 million per month. This allowed Vungle to accelerate its exponential growth to 4 billion video views per month over 1 billion devices, increasing its revenues in the period between 2014 and 2019 to more than $500 million per year, ultimately resulting in the sale of Vungle in 2019 to the Blackstone Equity Group for more than $750 million. 

Links

• Professor Bert De Reyck
• UCL School of Management
• UCL Faculty of Engineering Sciences
• Faculty of Engineering REF 2021

Image 

• Image credit: Shutterstock / A_B_C