Applications of Outlier and Anomaly Detection in Sponsored Search Advertising Campaigns (2011-12-16)
The project was an exploration of opportunities to quantitatively detect changes to the advertising marketplace that would affect your advertising performance. It seemed to be well received and garnered me honors status in the program.
Organizations using sponsored search advertising rapidly find their staff overwhelmed with the amount of quantitative data available to them. One area that is often overlooked is management of dynamic change in the online marketplace. This research attempts to provide a predictive model to determine when an ad group is likely to decline in profitability.
Correlation analysis shows that, at the ad group level, the advertising metric most predictive of change in 7-day profit margin is revenue-per-click (RPC). Additionally, the likelihood of a negative change in RPC predicting a negative change in 7-day profit margin can be as high as 76% when applying these methods to ad groups that have a high number of impressions. The likelihood of false positives is low (3%-7%) when the number of impressions is high, so applying these methods would likely yield an improvement in profit over ad-hoc analysis. The anomaly detection methods show considerably less effectiveness when applied to ad groups with fewer impressions and as such should not be used in an unsupervised manner without further research.
To see the full document, visit the University of Kansas Scholarworks repository:
Applications of Outlier and Anomaly Detection in Sponsored Search Advertising Campaigns