When it comes to digital advertising effectiveness, if something looks too good to be true, then it probably is…and it’s worth looking more deeply into it before you accept it as the truth.  For example, if you’re running an annual sales generation campaign and your video buys are outperforming expectations and industry averages, before rebooking the inventory, do a deep dive into the numbers, you’ll be amazed at what you might find. 

In digital performance media, amazing engagement rates are often the result of unidentified tracking errors at best, and fraud at worst. People ‘clicking’ is often not ‘people’ at all, nor engaged viewers more than likely it’s bots or clicks farms that hype the metrics to make unscientific clients and agencies over-rate media due to one or two fantastically good response measures. 

How machine learning can detect fraudulent patterns

Fraud detection capabilities within platforms such as IAS and Moat are great for verifying that your ad has been served and viewed in a browser and flagging certain fraudulent activities such as ad stacking and bot traffic. But with ad fraud costing advertisers almost £14bn a year, the investment going into new methods of fraud is far higher than the efforts that the likes of IAS and MOAT are able to provide. Using machine learning to complement the fraud reports within these tools will ensure you are being as diligent as possible in verifying the quality of your media buys. 

Detecting fraudulent activity should be heavily focused on identifying trends and patterns and digging into unusual patterns. This is where machine learning analysis comes into its own – especially since clients and media agencies know their media activity better than anyone else and is best placed to identify irregular patterns around campaign performance and media interactions. Leveraging internal data science resources can help add an extra layer of validation around the quality of inventory buys and minimise wasted media spends, especially for clients running large and sophisticated programmatic buying systems. 

Unsupervised machine learning techniques

Machine learning can be divided into 3 categories: supervised learning, unsupervised learning, and reinforcement learning. For spotting irregular patterns in an ongoing media and e-commerce programme, brands should use unsupervised learning techniques, whereby click rates, bounce rates and other site engagement metrics can be used to make observations about predetermined behaviours using predefined criteria.

Typically for fraud analysis, a carefully chosen range of variables should be agreed by agency and brand, with a desired range of volumes, interactions and sales rates; but, additionally, the algorithm can also choose valid ranges and populate heatmap charts visualisations for decision-making and reporting.

The benefit of this approach is that computers can analyse multiple measurement criteria and highlight unusual statistical patterns that may have escaped the attention of the media team. And they will do this much faster than a human could possibly make the calculations. And it is worth saying that this does not create a machine learning monster that runs media buying, but produces reports that can allow you to act whilst your campaign is running, rather than waiting for weeks after the campaign has finished for identifying these anomalies.

The key requirements to set up machine learning are to tag the media and site assets and create campaign logics that fit the brand’s needs and have access to a cloud computing instance that can handle multiple queries and datasets, such as Google Cloud Platform. And of course, you will need a data scientist to shape the inputs and outputs to ensure that they are in line with your business goals.  Implementing a strategy will give you the best chance to limit fraud and see the true picture of your advertising effectiveness.