Design of an ad targeting platform, matching ads to web pages. This system and its later modifications have since created millions of dollars of revenue for Seznam.cz.
Who, where, when
As the founder and leader of the research department at Seznam.cz, a major advertising company in the Czech Republic, I devised several web-related algorithms between 2008-2009. Together with my team we also implemented them; some directly into production code, some as prototypes, “blueprint” referential implementations.
The Sklik ad targeting platform is one of Seznam’s most profitable services. Its goal is to match queries and/or web pages to a database of ads/ad groups dynamically, in real time. The result is a “bidding market”, which you’re surely familiar with if you’ve ever used Google’s AdWords/AdSense. Sklik is the Seznam’s version of this popular platform.
My design and implementation consisted of several interacting subsystems: extracting information from web pages and ad campaign, collecting and extracting information about the user, and matching the two in a way that reconciles short-term profit (displaying highest bidding ads) with long-term profit (building the Sklik brand by offering highest quality matches/best expected CTR).
A major challenge in this system was data sparsity: the system typically doesn’t have enough knowledge about the query/page, or the user, to estimate the match quality (expected CTR) perfectly. The solution consisted of smoothing this data over larger sets of similar pages (similar theme, style, synonyms…) and/or users (similar profile). As a result, the system can display reasonable advertisements even for first-time users visiting newly indexed pages (unknown CTR), learn from user behaviour and improve with time.
Used tools, data, architecture
The detailed architecture and statistics of the Sklik ad targeting platform are not public.
The prototype implementation of the algorithms and pipelines was written in Python.
Ad Targeting / Machine Learning