dunnhumby’s Global Partner Summit: four things we learned about harnessing the power of AI for retail

Last October, leading industry experts and thought leaders gathered in Miami for the annual dunnhumby Global Partner Summit (GPS). In this—the latest in a series of posts reporting back from some of our keynote sessions at GPS—we’re looking at what dunnhumby’s Head of Research Data Science, Julie Sharrocks, had to say about ‘Harnessing the Power of AI for Retail’.

1. Different types of data science and AI serve different purposes
2. Machine Learning can be a supplement when data is limited
3. Generative AI models are powerful, but mostly when working with language
4. Generative AI is moving quickly

 
#1 Different types of data science and AI serve different purposes

"We use four different types of data science and AI," Julie explained. "Customer analytics, predictive science, machine learning, and generative AI. It’s not that one of them is better than the other," she added. "It’s about which is the right tool for the job."

Julie went on to explain that experimentation and proof of concept is really essential when working out what the most appropriate AI technique will be. The best approach is to think about the specific problem, and then assess the various different models that you have at your disposal.

 
#2 Machine Learning can be a supplement when data is limited

"It’s not always a given that our clients have an abundance of loyalty card data", Julie noted. "So we need to think more broadly than that. And that’s where machine learning can gives us boost."

Machine learning can augment data in quite a few different ways—enhancing the capabilities of data analytics and allowing us to extract more value from the information available.

Sometimes, it’s about using existing techniques in different ways. "We can use a machine learning technique which is commonly used for search term document retrieval, but instead of thinking about words in a document, we can apply it to products in a basket," Julie explained.

She went on to talk about how vector embedding algorithms like FastText can leverage Natural Language algorithms—in this case, products in a basket. "By understanding the context in which products are purchased,” Julie said, “we can create a ‘fingerprint’ for every product."

 
#3 Generative AI models are powerful, but mostly when working with language

AI, underpinned by Large Language Models (LLMs), is a huge help for language orientated tasks. "They’re great at helping us with things like naming and labelling science," Julie explained.

AI can help provide a name for a set of products that differentiate it from others, for instance. And accurate naming also helps with the navigation of customer decision trees. LLMs can also be really helpful for broader business tasks, like AI coding assistants and chat bots, Julie added.

 
#4 Generative AI is moving quickly

"Advancements are happening very fast. There’s lots of new development, lots of new models being created," said Julie.

Newer models can handle numerical values better than LLMs could previously. Many of them are scoring equivalently to top high school maths competitors. It opens up the possibility of where we can start to apply Generative AI to newer and trickier use cases.

While developments are happening at a breathtaking pace, it’s important to be clear on the use case and measure the benefit of adopting different types of data science and AI against your specific use case.

We’ll be back with more from GPS soon.