Machine learning algorithms operate at multiple levels across the platform, creating a fabric of intelligence backing the ad decisions on every page view with a focus on outcomes of these decisions such as clicks and conversion events. Some examples of Yieldbot Machine Learning are the real-time ad selection process, language modeling for intent extraction, forecasting for delivery prediction, and logistic regression analysis of performance over hundreds of dimensions.
Starting with how, when and where a visitor arrives at a site and continuing as visitors consume site content, their explicit session information feeds Yieldbot’s real-time intent analytics pipeline. This pipeline uses over one hundred real-time and historical data points to calculate a real-time Intent Profile™ – an index of scored keywords associated with the visitor’s session. As intent profiles change based on the click-paths, the high-performance lookup tables used during Matching are updated.
The core matching function combines policies, targeting rules, and real-time intent and performance data to determine applicable creative that can be served. High-performance lookup tables contain scored intent that is adjusted dynamically for every unique pageview. Decisions are made independently for every ad slot on each page view to account for the unique factors of every visitors context and site-session.
Backed by the powerful matching engine, real-time performance data, and optimization algorithms, the final selection of creative to be served is made. This decision is made independently for each ad slot on each pageview taking into account performance measured along multiple dimensions and considering the specific session as well analysis of past similar sessions. Our direct publisher-side integration also enables a rich choice of context-specific creative types.
Optimizing based on real-time performance feedback is done to maximize relevance. As ads are served engagement is measured and analyzed along multiple dimensions. Automated optimization algorithms operate on every page view to independently select the best option for the given circumstances.