AI models can detect complex patterns in customer behaviour, such as purchases and click and search sequences, earlier and more accurately. Luuk van Maasakkers, PhD candidate at Erasmus School of Economics, will defend his dissertation showing how this allows companies to predict much earlier whether a customer is likely to make a purchase.
On Thursday 23 April 2026, Van Maasakkers will defend his dissertation “Deep Learning Approaches for Customer Analytics”, supervised by promotors Dennis Fok and Bas Donkers. His work develops deeplearning algorithms inspired by generative transformer models, such as those used in ChatGPT, but adapted retail contexts. As he explains, ‘these AI models are well-known for predicting language, but with some adjustments, they can be trained to predict other types of sequences, like customer purchase behaviour.
Comparable to finishing a sentence
A key contribution of the dissertation lies in applying modern languagemodelling techniques to behavioural data from supermarkets and online platforms. The models learn patterns in shopping baskets, historical purchase sequences and click behaviour, enabling them to predict missing items or future actions. Van Maasakkers illustrates this by comparing a shopping basket to a sentence: ‘We took language models into retail contexts. If a customer buys pasta and pasta sauce, the model considers what the next product in the basket should be, such as Parmesan cheese, similar to how a language model considers what the next word in a sentence should be.’ When customers later shop online or scan items instore, the system can make personalised recommendations or offer relevant discounts.
The analysis extends beyond retail into settings such as insurance. The models examine which emails a customer clicks, whether they arrive via Google search, a partner website or social media, and how these interactions unfold over time. According to Van Maasakkers, such behavioural sequences can help identify early purchase intention: ‘When someone shows search behaviour suggesting they are really looking for insurance, the system can identify this in time and the company can adjust its marketing actions accordingly.’
More accurate than traditional statistical approaches
Across these applications, the deeplearning models outperform traditional statistical approaches. As Van Maasakkers explains, ‘If I buy apples and bananas, and my neighbour buys apples, bananas, and pears, the old method recommends pears to me.’ The new models capture far richer patterns: ‘We see that with these new techniques we can can capture more flexible, diverse patterns’, adding that ‘the more difficult the task, the greater the difference between an AI model and a traditional approach’.
This is because traditional methods rely on simplifying assumptions about customer behaviour, whereas deeplearning models learn patterns directly from data. Van Maasakkers puts it simply: ‘In traditional methods, we make many assumptions about reality. With machinelearning models, we let the algorithm learn all patterns itself. When trained well, it has more flexibility to learn different types of patterns.’
Insights for marketeers
The findings offer practical implications. Retailers and ecommerce firms may use these insights to refine or redesign recommendation systems, enabling more accurate predictions of complementary products. Companies analysing customer click and search behaviour could identify potential buyers earlier and target them more effectively. The research also shows how such models can be made scalable and operational for largescale customer data.
Van Maasakker’s research and its applications also benefits consumers. They receive better personalised offers and recommendations, and with these techniques, shopping gradually becomes easier and requires fewer actions.
And perhaps something could be added/acknowledged about the fact that these kinds of applications may raise privacy concerns. Furthermore, he wants to acknowledge that these kinds of applications may raise privacy concerns, explaining: ‘In one of our studies, we show that with minimal purchase data that doesn’t contain detailed personal information, we can already generate good personalised offers using these models.’
- More information
For further questions, please contact Ronald de Groot, Media & Public Relations Officer at Erasmus School of Economics: rdegroot@ese.eur.nl, +31 6 53 641 846.
