Kantar uses Azure Machine Learning to help streaming services drive market growth

To stream or not to stream?

When it’s finally time to sit on the couch and turn on the TV, you’ve got a lot of options. Netflix, Hulu, Disney+, HBO Max—the list goes on. Different streaming services offer diverse content and distinct tiers of service, with their own niches in the market. With only so much time to watch TV and a budget to maintain, you have decisions to make.

But what’s driving those decisions? What causes someone to unsubscribe or switch to a new service? Or in the back-office argot at Hulu or ESPN+, why do they churn? Kantar Group, one of the world’s leading data, insights, and consulting companies, is working with the major players in the industry to not only answer those questions but also help its clients anticipate customer behaviors, make the right decisions to retain their audiences, and predict future outcomes and growth. 

Propensity to churn

Worldpanel, a division of internationally based Kantar Group, having launched a new business unit, Entertainment on Demand, includes as part of its offering a customer loyalty analysis solution to help entertainment clients plan for growth. It uses Microsoft Azure Machine Learning to support a Propensity to Churn platform that effectively analyzes all its data to profile customers who are more likely to cancel their subscriptions or switch to a new service. With this capability, it now can provide a quantifiable representation of churn likelihood within complex user-defined groups.

While Entertainment on Demand is only 1.5 years old, the Worldpanel division has been collecting longitudinal customer behavior data for more than 50 years. The Entertainment on Demand team wanted to go beyond simple binary insights into retention—a customer either stays with a service or leaves. Kantar could identify who might churn and who might not, but it couldn’t quantify the probability of churning for any individual customer. Kantar leadership understood that the company needed more computing capacity and more machine learning power to turn its available data into the kind of predictive insights its clients needed.

“We’ve got such a wealth of data,” says Jennifer Chan, Head of Client Service for North America, Entertainment on Demand Business Unit, Worldpanel Division at Kantar Group. “If we’re not using everything to meet client demands through predictive modeling, we’re not doing our clients or our company justice.”

The right solution

Although Kantar had tried out machine learning platforms with various companies that provide BI solutions, it wanted everything—data storage, data cleaning, machine learning and insights, and visualization all within a single solution. The company needed a more robust, cloud-based system to scale its computing power and analyze its bulk of data. After a positive experience using Microsoft Power BI to visualize data and help deliver faster reporting to clients without having to run the data and paste it into PowerPoint presentations, Kantar decided to look to Azure for both its cloud infrastructure and machine learning solutions.

In addition to using Azure Machine Learning to train its AI models, the Kantar data science team chose multiple Azure platform as a service (PaaS) components for Propensity to Churn. The platform uses several different types of cloud storage technologies, from Azure Data Lake Storage to Azure Blob Storage, and more recently, Kantar has started experimenting with Azure Data Factory to store and manage its Azure Machine Learning notebooks. It also uses Azure Analysis Services as the analytics engine for its data models, Azure SQL Database for intelligent SQL in the cloud, and Azure Cognitive Services to deploy high-quality AI models as APIs. By using Azure PaaS resources to simplify and streamline data storage, processing, and analytics, Kantar can automate simple tasks, let operations teams run relatively sophisticated programs, and focus its highest-value engineers and data scientists on its most critical and complex workflows.

“You can see how heavily invested we are in the ecosystem of Azure,” says Hideo Ogawa, Head of Data Science at Kantar Group. “With this infrastructure and our data in Azure Data Factory, we will be able to simply pass information to our operations team, and they don’t have to be data scientists to run the modeling programs.”

The models

In the past, it took the Kantar data science team a great deal of time and effort to create its models. The complicated manual processes slowed engineers down and made it hard to keep up with other workflows such as system maintenance and random ad hoc projects. They wanted an interconnected solution with flexible tuning so that they could build a more robust machine learning pipeline, scale up for reporting, and complete more projects. 

Kantar uses 10 models for its Propensity to Churn platform and takes advantage of automated machine learning in Azure Machine Learning to accelerate model training and validation. With the ability to automatically scale computing power as needed, Ogawa and his team have streamlined model production and increased model accuracy by running more models simultaneously and aggregating models to generate more granular results. Working with up to 5 gigabytes of data at a time, the team can generate behavioral clusters with what Ogawa calls “tremendously high” precision scores of 85 percent or more.

“By using Azure Machine Learning, we can now train 10 models in the time we used to spend on 3,” adds Ogawa, “and the sum of the additional models generates more accuracy for our clients in less time and with less effort.”

When building models, the team looks at multiple parameters such as customers’ reasons for subscribing, usage behaviors, Net Promoter Scores, tenure, demographics, subscription-renewal type, purchase journey, satisfaction drivers, and content titles. The Propensity to Churn platform uses the models to process all that data and deliver critical insights that Kantar clients use to make important marketing, programming, and other business decisions. For example, Kantar has found that younger customers have a higher propensity to churn.

But it’s not as simple as throwing data into a model and capturing insights. Kantar used Azure responsible AI—a framework for safer, fairer, more transparent machine learning models—to improve its data and model quality. The company uses interpretability capabilities in responsible AI to identify drivers that are high indicators for bias in its data, making it easier to comprehend results. After it used interpretability to identify that certain brand names were high influencers, the Kantar team began working within brands individually to generate more accurate, more useful results.

“With Azure Machine Learning, we’re bringing entirely new perspectives to our clients,” says Chan. “That helps them generate accurate, unbiased insights and make better-informed decisions based on past customer behavior.”

The future of streaming

Streaming services get more from Kantar than customer profiles and the probability of churning. By drawing on the insights it provides, clients can boost customer loyalty, and they can make data-driven decisions that help them to keep customers engaged and minimize the risk of service cancellation or disuse. Clients use the predictive modeling results to better determine future growth targets such as propensity to acquire or propensity to spend and act with confidence—minimizing loss, reducing costs, and ultimately increasing revenue. And with a stronger focus on their customers’ experiences and preferences, the streaming services help their audiences get more out of the content they watch.

“There are a lot of vendors out there who can just review and evaluate what’s happened,” says Chan. “But the client demand is to predict future outcomes and future growth, and with an Azure ecosystem, we can do that.”

Building a new platform based on Azure Machine Learning and its automation capabilities has not only helped Kantar quickly deliver highly accurate critical analyses and satisfy its clients, but it also replaced complex, time-consuming, manually written Python code. “Our data science team estimates that by using the whole Azure ecosystem, they save about three days a month on model creation and approximately three weeks a quarter on reporting,” says Ogawa.

Kantar has made its Propensity to Churn modeling platform into a valuable asset to the Entertainment on Demand business unit, both in terms of delivering actionable insights and generating at least a 20 percent incremental revenue uplift to subscriptions and projects.

Find out more about the Entertainment on Demand service here.

Find out more about Kantar on Twitter and LinkedIn.