A group of Microsoft Research scientists published a paper that attempts to build a benchmark to analyze the presence of "interference" in sponsored search advertising. The original, published in 2020, and resurfaced in July 2022, attempts to call attention to how each advertisement on a page influences the others. The research has become more important as the industry moves away from cookie-based tracking in consumer browsers and looks to find other alternatives to target relevant ads.
Casual inference assists in making data-driven decisions for online advertising. The model means to infer cause and effect relationships with any type of data is complicated.
Then there’s interference. Some experts believe when built into statistical models, interference can compute rank scores for each ad to help optimize the final layout of each search page. In other words, each ad on a page influences the other, especially as the industry moves away from cookie-based tracking in consumer browsers.
“A proper understanding of the interference issue in relation to causal inference directly impacts engineering of more purposeful interventions and design of more effective A/B testing for ad placement,” wrote several Microsoft Research analysts in a paper titled Casual Inference in the Presence of Interference Sponsored Search Advertising. “Modeling what we are after in the context of sponsored search advertising is closer to the causal framework for modeling interference in social networks.”
The group attempts to set new benchmarks for search advertising by analyzing the challenges of interference among ads in what they believe is the first paper on the subject. The hope is to create a benchmark for future work in search advertising that goes “beyond the classical iid assumption,” which states that random variables are independent and identically distributed.
What does independent and identically distributed (iid) mean for search advertising? All the variables provide the same kind of information independently of each other.
Using graphical models, the team assume a causal structure that uses various sources of interference to find statistically significant interference effects among ads.
The work points to quantifying the cause-effect relationships between a variable or causal effect with an outcome using experimental or observational data.
“Causal inference uses assumptions in causal models to link the observed data distribution to the distribution over counterfactual random variables,” researchers wrote.
The model uses three main assumptions: “Consistency,” “Conditional ignorability,” and “Positivity.”
The paper details sources of interference in ad placements, interference effects among ads, identification of assumptions, elimination of causal effects, structure and more.
It turns out that each ad serving up in close proximity influenced the others. The paper focused on impressions of ads on the search result page, and marginalized the ads involved in the search engine auction.
“Incorporating the knowledge on how exactly the auction optimizer works on the entire set of candidate ads is important in determining the optimal layouts in presence of interference,” the Microsoft Research team wrote. “We further restricted our attention to auctions that only yield two blocks on the final pageview.”