Predictive Programmatic Classification Model to Improve Ad-Campaign Click Through Rate

Abstract

Programmatic advertising is a component of digital advertising where software is used to buy digital ad space in real time connecting advertiser to a specific consumer. Gaining more clicks and improving Click Through Rate (CTR) is considered as a goal for many of the programmatic advertising campaigns. The average CTR across the internet hovers around 0.2% but generally fluctuates somewhere between 0.1%–0.3% making clicks a rare event. Improving the CTR helps to manage the campaign budget efficiently. So, it’s very important to get high CTR values for the campaign to be effective. In this paper we present a classification model that will predict whether the user will click an impression or not. Most of the state-of-the-art works in this area are focused on learning the feature interactions and capturing the user interest evolving process. As majority of the features in ad-techspace are categorical type and with high cardinality, it’s very important to encode the features to best relate it with the class label and reduce the dimensionality. Less attention is being paid to the encoding part by most of the state-of-the-art works. The proposed method uses a custom feature encoding which is underpinned on class label distribution. A hybrid approach comprising class weight, stratified cross validation and probability thresholding is used for dealing with class imbalance. f1-score is set as the model evaluation metric and proposed encoding scheme showed a 68.55% test f1-score whereas the classical one hot encoding showed just 60.92% f1-score.