Data Science Investment Counter —Funding raised by UK data science companies in 2018.
£ 5.640 Million

How Yahoo! Singles out Shoddy Ads

Yahoo! is using machine learning to identify low-quality online ads and understand why they fail to engage users.

Speaking at  London Data Fest, Yahoo!’s Director of Research Mounia Lalmas explained how her team developed a model accounting for why some advertisements work and others do not.

Most of Yahoo!’s efforts, Lalmas said, are focused on native advertising, especially on mobile platforms. Lalmas explained that, to asses an ad’s success, the focus is on the so-called “post-click experience”, rather than on the sheer amount of clicks it gets.

My job is to make users happy while [advertisers] make money.” she said.”The secret is to make this work for users. And if users’ experience is negative they won’t come back.”

The main metric Yahoo! resorts to in order to evaluate users’ experiences is their “dwell time”— that is, how long they stay on an ad page and how they interact with it.

Lalmas’s team uses machine learning on interaction data to identify what traits are more likely to be present in successful ads. In this way, they manage to tell advertisers what features should be tweaked—from the number of words, to colours, to imagery and structure, to sentiment analysis— to maximise an ad’s effectiveness.

Machine learning techniques are also employed to understand when user engagement is genuine rather than due to accidental clicks.

The model, Lalmas said, is helping advertisers create higher-quality ads; she underlined, though, that the learning process is continuous.

“The model has to change over time, because, for instance, the level of engagement we witness usually can shift radically when Christmas happens,” she explained. ” That is why we retrain the model regularly.”

Image via


Co-working space and blog dedicated to all things data science.

Subscribe to our newsletter