By Eric Shiffman, VP, Product Marketing
For every impression, there are many creative format experiences that can be served, but how do we know which is best? Our whitepaper, “Elevating Creative Performance with Predictive Creative and Artificial Intelligence (AI)” solves this.
Earlier in our blog series, we dove into the first 2 components, formats and training data. Today we’ll cover the final 3 components: modeling, machine learning operations, and activation.
Creating Data Modeling and Machine Learning Processes for Predictive Creatives on Activation
Once we’ve processed, transformed, and loaded all these data into our big data processing tools, we kick off the first stage on our machine learning pipelines. For every individual historical ad event we’ve collected, we need to assign a probability that a user takes some action we care about. This could be an attentive view, an ad engagement, a click, a video complete, or anything else… because the interaction data we collect are so prevalent and granular, they can be used to predict other high value user actions.
The machine learning challenge here is to train a model that can reliably use these signals to predict a given action and use those to assign a probability of each type of action to every event (this is known as scoring). We assign probabilities of every historical event, refreshed daily, using neural networks.
Once every historical event has a probability, we kick off the next stage of the machine learning pipeline and use those probabilities to extract and rank the most predictive feature combinations from highest to lowest (these are the massive permutations of millions of possible combinations of different values of advertiser, context, and user signals which we discuss in the whitepaper) that can be used in real-time to predict which format to serve under different conditions. Yieldmo has hundreds of models running in parallel, and actionable within our exchange.
Once we have this set or predictions for each feature combination, we surface them to the ad server so that for every ad request, we can figure out which ad format to render in real time, all within a few milliseconds. In our case, in our frictionless setup, all the creatives are hosted by the DSP to reduce latency and streamline implementation.
Finally, to ensure that advertisers are benefiting from the most recent data about how users are engaging with their creative formats, user interactions are logged and rescored every time an ad is served, with feature combinations refreshed daily.
This process is incredibly fast, creating 7 million unique predictions every second. In the end, our solution results in about ½ a trillion events being optimized every day.
In our final blog of the series, we’ll share how after 2+ years developing this end-to-end workflow, we put it into action and drove real results with partners like Groupm.
Can’t wait to learn more? Download our whitepaper, Elevating Creative Performance with Predictive Creative and Artificial Intelligence (AI).
