Our client is a leading digital media company that is transforming mobile advertising through data-driven predictive solutions.
To be competitive and deliver the best value to their clients, they needed reliable predictions of win probability as well as the probability of click-through of each digital ad in their inventory to determine the optimal price for real-time bidding.
- Tens of millions of data points to analyze every hour
- Attributes with many levels, amounting to more than one million points
- Clicks are very rare events – typically one for every 3,000 impressions.
- The raw data consisted of ad impression attributes, publisher attributes, category restrictions, and user profiles
- Performed feature engineering using feature hashing and CTR encoding to handle columns with large number of categories
- Reduced the number of data points by ten-fold without loss of prediction accuracy through case sampling
- Developed a predictive model to predict the win and click-through probability of each ad impression. Validated the models on sub-samples of various mobile platforms. One of the chief insights obtained was that the top 20% of the impressions capture 43% of the clicks
- Built a robust solution which can accommodate constraints like category restrictions and new websites for which historical data is not available