TikTok team lead explains how it makes predictions that benefit customers
The possibility of a U.S. ban on TikTok following the passing of a controversial bill has started a conversation that usually revolves around what it will mean for users. These include the approximately 170 million Americans who use the short-form video sharing service. Those like Sarah Norman, however, are focused on another group that may have more at stake than simply losing a source of entertainment: TikTok’s paying customers.
As TikTok’s team lead for enterprise, Norman and her team apply data science to better understand how advertising performs on the platform. The goal is to prove that, beyond merely getting in front of a large audience, ads on TikTok will move the needle for brands. Watching an ad might lead to someone contacting the brand directly, for instance, or making a purchase.
Speaking at the recent All Access: Predictive CX virtual event with CX Network, Norman provided some insight into how the right approach to modelling data could benefit customers, as well as some of the pitfalls to avoid.
TikTok
Though TikTok has amassed a large amount of data to work with, for example, Norman pointed out that it’s surprisingly easy to overlook potential opportunities to enhance your products and services. She pointed to Meta CEO Mark Zuckerburg’s admission that he did not foresee the rise of short-term video – the kind of content that helped TikTok to thrive.
“He has more social media data than anyone. He certainly has some of the world’s best data people and engineers and predictive models,” she said. “(That data) did not do a good job.”
Halve the customer data to test predictive accuracy
The issue, Norman said, is that even the best predictive models tend to be based on historical data. What happened in the past will not always tell you what’s going to happen in the future.
This is what makes testing so important. Norman recommended splitting the volume of customer data you’re using in half, and applying a predictive model to only one of those halves. Using a sample that spans the previous three years, she suggested brands could compare the two data sets to see how well a predictive model identified trends in customer behavior.
Predictive modelling in CX can also be compromised by a lack of data quality. Norman used the example of a government trying to determine the best places to build housing. This could be based on demographic data, but that assumes everyone in a given population has filled out their census form.
“You can adjust the model and weight those groups differently, but as humans, we tend to trust our data sets, especially one from the government or a really big company,” she said. “We don’t tend to question whether that data was accurately collected.”
Make it okay to call out data quality issues
The hopes brands put around predictive modeling for CX purposes can inhibit those who might otherwise speak out, Norman added. That’s why it’s important leaders create an environment where anyone can call out the limitations in their data.
As an example, Norman referenced a public sector project she worked on to determine why people weren’t signing up for resources they could receive amid the COVID-19 pandemic. She said she believed it was due to the web site being overly complex, but she had to admit nothing was certain unless people could be surveyed on the subject.
“You really have to be honest with people. Your analysis is your currency. Your analysis is your value,” she said. “The second you give a data analysis that is really incorrect without being honest up front about your limitations, you lose all your credibility.”
Try the triangle method
One way to avoid some of these missteps is to apply what Norman described as the “triangle” method. This is where you use data modelling to make prediction in short, medium, and long-term time frames. If the data points to the same answer all three times, you’re probably on the right track, she said.
At the same time, predictive modelling needs to be balanced with human judgement. Norman pointed to the luxury beauty sector, where incorporating cannabis into products is becoming more mainstream. While some companies might have thought such a move would diminish their brand’s image, some took an educated guess that cannabis would make products seem even more luxe.
“That human component won over the predictive model in that case,” she said. “Now some brands have a head start.”
Confront the data gatekeepers
In some cases, CX leaders may lack the data they need to build predictive models because they are housed in organizational silos where teams act as gate-keepers. Norman said it’s a matter of knowing when to pick your battles, but asking politely and explaining the purpose can help. On other occasions, she admitted to going over someone’s head to make her business case.
There can be legitimate reasons why data can’t be used to feed a model. “I would never be able to access someone’s personal data at TikTok — nor should I be,” she said. “Just because you own it doesn’t mean there aren’t rules around it.”
Become a better data storyteller
The key, Norman said, is not only to build predictive models but learn how to become a more effective data storyteller. This starts by showing the stakeholder you’re trying to convince to use predictive modeling that you understand what they care about, and that making better use of data will help them achieve their goals.
Good data storytellers also learn now to drown their audience with a hundred different pieces of information.
“You need to curate,” she said. “Get it down to one chart that shows them what the data could do.”
Make it short, compelling and easy to consume, in other words. Just like the content on TikTok.
