Enhancing Online Advertising Key Performance Indicators Monitoring: A Cost-Effective and Automated Anomaly Detection Framework

This thesis investigates anomaly detection, specifically in the context of online advertising. The main obstacle was the lack of labelled data, which required an inventive solution to properly evaluate our framework. As a result, feedback from account managers and analysts was crucial in perfecting our anomaly detection system. The framework presented here covers the company’s two main businesses: Affil- iate Marketing and Media Advertising. The former, which is aimed at advertisers and can monitor around 1,000 advertisers and a million orders a day, and the latter, which is aimed at media advertising and monitors KPIs for publishers, DSPs and AD spots in more than 100 countries. The results were impressive, with numer- ous commission savings for advertisers amounting to tens of thousands of dollars. Account managers also gained actionable insights that led to better deals with publishers. Remarkably, the entire advertising flavour operates at a cost of less than $5 per day, including database queries, storage and framework maintenance. Similarly, the media flavour, which is even more extensive, remains cost effective at less than $1 a day. Although we are awaiting feedback from the media team on anomaly detec- tion, early indications suggest a positive reception, with the framework successfully identifying fluctuations. In conclusion, the data science team has implemented a cost-effective, auto- mated and insightful anomaly detection framework that has become an integral part of our company’s operations. This innovation ensures that our internal teams now have an automated, efficient method of monitoring customer performance, leading to more informed decisions and improved customer relationships. As we move forward, the data science team is committed to continually refining the frame- work and exploring opportunities to extend it to new businesses in this ever-evolving field.