Anthony Kilili: ‘Stay true to what’s right for the Customer.’
dunnhumby’s Head of Data Science, Media and Customer Engagement, talks personalised recommendations, behavioural predictions, and why nothing matters more than the Customer experience.
One of dunnhumby’s key products is Recommenders, which allows Retailers and Brands to deliver personalised product suggestions to shoppers across a wide range of channels. A lot of the work that my team does is centred around finding the relationships between products and Customers to ensure that the recommendations we’re delivering are relevant, accurate and effective.
Recommenders can take a number of different forms – complements, substitutions, items that shoppers might have forgotten to add to their cart, and more. Generating them means that we need to have a complete, 360° view of the Customer, as well as predictive models that can accurately forecast their future needs. That’s something that we achieve through a combination of machine learning/artificial intelligence, data, and a huge amount of computing power.
We use over 15,000 datapoints for an average Customer. Obviously, these are all unbiased and anonymised, but they give us insight into what shoppers are buying, when, how much, which pack types they prefer, and much more. This allows us to build deep Customer and product understanding, which we can then analyse in order to determine their relevance to one another.
It’s all about accurate prediction. We use machine learning algorithms to make forecasts about how likely a Customer is to buy something within a certain time period – what’s known as a propensity model. When you cross-reference that against which items are on promotion, you end up with something really powerful in terms of a recommendation: “here’s something we think you’ll like, and it’s also on sale”.
We help consumer-packaged goods (CPG) Brands find their audiences. They have a product they want to market, and doing that successfully means that they need to find the “best” target Customers for a campaign. We can help them, because we’re able to accurately predict the most relevant shoppers for that product. It might be people who’ve bought that item before, it could be new acquisitions; what’s most important is that the product is relevant to them. The added benefit is that we can see exactly how effective those predictions are using the same loyalty data we built them with. If those shoppers buy the item, the prediction was right. If they don’t, we can learn why not and refine it further.
Relevancy isn’t just about effective marketing, it’s about the quality of the Customer experience. The shopper must benefit from sharing their data with the Retailer. So it’s never about selling a promotional listing for the highest price, or letting Brands advertise just because they’re willing to pay. If you don’t approach this from the point of view of doing what’s right for the shopper, you’ll only end up wasting money and damaging their opinions of your brand.
Context is another key aspect here. No-one likes being shown an advert for something they’ve already purchased and no longer need, and so you need to understand when a Customer is genuinely likely to buy something. That applies to their behaviours beyond just purchasing frequency, too. If your data tells you that someone is visiting a store on a hot day on their way back from jogging, it might not be the best time to send them an offer for soup!
The same logic applies for channels. It’s no good targeting shoppers via email if that’s not a medium they want to engage with. It might be that they prefer paper coupons, or that they like the convenience of an app. The important thing is being able to understand them so well that you can communicate with them as they want, not how you’d like to.
It won’t be too long before we’re dealing with petabytes of data. We process terabytes of shopper data today, but the addition of elements like video and image data will push us beyond that relatively soon. These will help us take recommender science to the next level. We’re also focused on ensuring minimal lag between data capture and processing. Recommendations should update the moment that you make a purchase.
Recommenders played a key role at the start of the pandemic. When supply chains were disrupted and many items were in short supply, the Substitutions side of the product was able to suggest alternatives. We know how well that worked, because we track “refusal” rates for substitute items. These dropped to extremely low levels, so it’s good to know that we were able to help people maintain some degree of normality in spite of everything that was going on.
The future of personalisation is really exciting. It all depends on our ability to think creatively about applications. What if we could build entire baskets for Customers before they start shopping? What if we plug into connected devices and offer recommendations based on what they currently have in their fridges or freezers? Could we suggest new recipes based on items and quantities they’ve already purchased, helping them tackle food waste and get better value for money? The potential is all there – we just need to stay focused on how the Customer benefits.