By Antoinette Siu • January 2, 2024 • 4 min read •
Ivy Liu
This article is a WTF explainer, in which we break down media and marketing’s most confusing terms. More from the series →
As the agency world focuses more intently on data standardization and metrics, companies are turning to anomaly detection as a way of finding outliers and abnormalities in data points.
This data analysis process can help agencies and their partners (be it clients or ad-tech contractors) in identifying suspicious patterns, monitor compliance for clients and ideally serve as a preventative measure against harmful data — sort of like how banks use it to target fraud.
Here’s a look at how this strategy works and why it’s becoming more important to integrate as privacy needs increase and standardization efforts evolve.
What is it?
Anomaly detection examines data points to detect outliers and other unusual patterns in the data. Businesses use it to prevent fraud, breaches, exposing sensitive information while also monitoring for other abnormal activity. The method behind it may not be new, but it is gaining attention as more data becomes available — it offers a more efficient way to monitor data compared to manual tracking.
How does it work? What is an example?
Anomaly detection combs through all data and looks for any shifts that have occurred, be it in the data type or data volume that is set by defined parameters. The goal is to detect any irregularities in the data where something is broken, such as an increase or decrease in spending at different rates. These could be also activities and events that are considered outside of the defined normal behavior.
“You’re looking for data points, and these are all very tailored to what individual brands are experiencing,” Michael Neveu, senior director of machine learning and AI solutions for Media.Monks, recently told Digiday.
Why are people talking about it now?
The use of AI and machine learning can lead to more of these anomalies. For instance, tools like Bard and OpenAI can produce “lies” or make mistakes, explained David DiCamillo, CTO of Code and Theory. Those platforms tend to take data from various sources to spit answers back to humans, which is why they also add disclaimers reminding people to check sources and verify the information. These are also called hallucinations that are considered to be deviations from the normal operations in AI systems, according to DiCamillo.
“The brain of the AI is just the data that it’s programmed with, so if there are anomalies in the data, then we can expect to get some level of hallucination,” DiCamillo said.
Why is it important?
Anomaly detection is part of a toolset that can help agencies and partners improve processes like forecasting, inventory management and customer experience, said Ram Singh, chief performance media officer at Crossmedia.
“Doing this right has further downstream benefits,” Singh said. “Tight performance management through such advanced detection and standardization helps improve forecasting, revenue and labor planning, inventory management, customer experience … [using these models] to detect problems quickly [allows us to] pass the savings back to the client — or use the savings to elevate client service and fund more advanced solutions to push performance beyond goals.”
Brands are also interested in anomaly detection tools to protect brand safety and monitor data and privacy compliance metrics. This strategy helps them find significant changes and respond quickly as the volume of data and sources grow.
DiCamillo added that the development of AI will require constant monitoring for irregular data. “We like to tell clients that implementing AI is sort of like welcoming a puppy into your house — one that will never grow up, will need constant oversight and will never be potty trained,” he said. “Your AI tools need constant maintenance and oversight in order to perform as expected.”
How are agencies leveraging anomaly detection?
They are using it for data standardization practices and leveraging machine learning to build models for detecting and responding to problems quickly. This ultimately helps with media performance, and “anomaly detection is essential to standardization processes,” Singh said.
There are also a number of AI tools, such as Fiddler and Arize AI, that help analyze data and monitor AI-produced information. For now, there is no single software that detects all these data outliers, so it will still require human oversight, DiCamillo explained. In the meantime, agencies will have to continue relying on their analysts and user feedback to flag anomalies, as well.
“We recommend regular system audits, activity logging and the maintenance of detailed audit trails,” DiCamillo said.
These can help surface anomalies and allow agencies to make adjustments and retrain their models and algorithms — especially as AI’s influence over data gathering and analysis takes shape in the agency business.
https://digiday.com/?p=529871
