Data Driven Recommendations for Display Advertising

Advertisers running display advertising campaigns often request actionable recommendations for booking the most effective new ad campaigns and improving the performance of ongoing campaigns. Typically, the recommendations desired by advertisers fall into two broad categories: improved performance in terms of metrics like CTR, CPC, and CPA; and increased reach, which is the number of unique users exposed to the campaign. Account managers provide recommendations to advertisers based on their personal intuition of the advertisers’ needs. This approach is not scalable and recommendations are often not consistent across account managers and advertisers. We developed a data-driven approach that leverages historical ad campaign information and granular user data coupled with the advertiser’s current campaign objectives to make effective recommendations. This paper presents the following key results: 1. A novel application of the PLSI algorithm for effectively identifying neighbors of ad campaigns. 2. Application of large-scale collaborative filtering methods for making recommendations to optimize ad campaigns. 3. Design of a complementary user segments algorithm to significantly increase the reach of ad campaigns, while maintaining or improving performance. The key advantages of this method of producing recommendations are scale: even small advertisers who do not have dedicated account managers can leverage them, and novelty: mining historical campaign and granular user level interaction data enables discovery of non-obvious recommendations.