Buying or just browsing?
Conversion latency prediction In Ecommerce digital advertising, the purchase value, made after a click on an ad, is considered the “return on ad spend” (ROAS) and is the key performance metric used to evaluate the
efficiency of an ad campaign. “Conversion tracking” is the process of collecting and attributing online purchases to clicks on ads, of a specific campaign, on a specific day.
In this talk we will share the challenges of collecting the data for this prediction model and the infrastructure that enables us to experiment at scale. We will describe how the data is represented as an atypical time series of updates and the data augmentation we apply in order to feed a recurrent model component. In addition to time series data, two methods are used to create distributed representations of advertiser campaigns. First, a custom language model is trained as a supervised classifier on an alternate prediction task and embedding vectors are extracted. Finally, an entity embedding of the advertiser campaign structure graph is used for an additional feature vector representation.