A revolution in online grocery retailing
Unilever is developing new technology that prompts consumers to buy seemingly unconnected yet relevant and suitable products.
The system works by recording the purchasing decisions of previous users to build up a complex model that can predict the most likely combinations of a person’s shopping e-basket.
Ming Li, Unilever R&D, who helped develop the system says, “If someone with similar shopping habits to you buys certain items of fish, ice cream and deodorant, when you come along with your basket containing the same fish and deodorant, the system will recommend ice cream.”
With book, music and video websites, models are based on what other customers have bought together and how highly they rated the product in feedback. And some online grocery retailers use logic-based rules to prompt a consumer to buy, say, biscuits when they select cheese. But our tool is significantly more sophisticated.
Ben Dias, Senior Research Scientist explains, “Grocery shopping is more complex due to the huge variety of categories, the lack of explicit feedback on preferences and the fact that shoppers frequently buy many of the same items. Also each product has a different rate of consumption. In future, the system will take that into account and, in time, will prompt based on typical usage.”
The performance of early prototypes was limited due to their inability to explore transitive associations between products that have never been co-purchased. We overcame this by incorporating a basket-based random walk algorithm in a model similar to Google’s PageRank link analysis. This calculates an importance score by also taking into account the consumer’s current shopping behaviour.
A trial with the leading Swiss online supermarket www.LeShop.ch (subsidiary of the country’s number one retailer MIGROS) led to outstanding revenue growth. According to Dominique Locher, Marketing & Sales Director, “Consumers using the tool spent an average of six per cent more per visit.”
We are one of the few research organisations to be working on recommender algorithms specifically for the grocery sector. Ming Li sums up, “We continue to evolve the system through keeping track of personal consumption rates, suggesting things the shopper may have forgotten and including information such as exact time of purchase. The main challenge is to make this technology quick enough so it can generate recommendations within the time it takes a web page to refresh.”
Unilever R&D also collaborated with Professor Paulo Lisboa and his team at Liverpool John Moores University on this project.