PROFILE: Josh Redelinghuys wants you to know your wine

Shopping for wine before the weekend? You’ll need good luck if you plan to rely on a retailer’s recommendations about what you might enjoy.

“The recommendations are simple and somewhat arbitrary. And for people who don’t drink wine and who aren’t willing to spend money to experiment, they are quite impractical,” says self-confessed wine nerd Josh Redelinghuys.

There are plenty of recommender apps that can help. But they treat South African wines as an afterthought, he says. Vivino, for example, lists only a few hundred wines from a country that has about 2,700 wineries. Nevertheless, it has a following: in March, Gauteng hosted the first Vivino event in Africa, attracting 96 users who tasted wines from nine producers.

When it came to deciding on a topic for his master’s thesis in data science, Redelinghuys knew he could do better. The product of his 18-month “passion project” at the University of Cape Town (UCT) is SoMLier (it’s “sommelier”, or wine waiter, with the abbreviation for “machine learning” in the middle), an app devoted to South African wines.

For now, it’s a proof of concept built on the Shiny platform, and without the resources to develop it into a mobile app, Redelinghuys has taken a job as a data scientist with consulting company Accenture.

But SoMLier is usable, and Redelinghuys says it generates “accurate, high-quality recommendations … and can assist consumers in discovering wines they will likely enjoy and understanding their preferences”.

The 24-year-old’s paper on SoMLier has been published in the Journal of Wine Research, and marks the high point of a wine journey that began when his oenophile father took him for a week-long wine-tasting tour of the Robertson area after he finished matric.

“Until that point, I had a very teenage understanding of wine. The tour was eye-opening because I saw how diverse it is, how delicious it is. It clicked in that moment,” he says. “Sampling wines at the places they had been grown and produced and nurtured became so much fun to me.”

It’s a habit he still exercises most weekends, and he values the chance to taste wines before buying them. Buying without tasting “forces novice consumers to purchase the wine based on other factors such as the design of the bottle, its price, awards it may have won and descriptions on the back label,” he says in his paper, supervised by senior lecturer Şebnem Er.

His research also found that the recommendations provided by 58% of popular wine retailers (the rest don’t make recommendations) “lack insight and can often seem meaningless to the consumer”.

Redelinghuys began building SoMLier with the assistance of Vivino and wine.co.za, which allowed him to use databases containing ratings and reviews for 1,640 wines from 115 farms and 37 regions.

Many equations later, he has an app that combines three approaches to generate recommendations.

The first is item-based collaborative filtering, which uses a consumer’s ratings of wines they have drunk to generate suggestions for others they might like. Then it’s on to a content-based approach, which uses wine properties such as acidity, tannins and region of origin to measure similarity to wines a consumer has enjoyed. Finally, algebra predicts the ratings consumers would award recommended wines as a way of refining the app’s list of suggestions even further.

SoMLier recommendations begin with text such as: “You have rated five wines from three different regions. The average price of your rated wines is R200, and you most commonly rate white wines from the Western Cape. Viogniers are your most favoured as you award them an average rating of 4.5. Pinot grigios are less enjoyed with an average rating of 2.5. Your top 20 wine recommendations come from 10 different regions. The average price of your recommended wines is R206.50, where mainly white wines from the Western Cape are suggested. Blends and cabernet francs are predicted to be the most enjoyed and some recommended viogniers are also worth trying.”

“I know I would benefit myself if my little humble application grew legs. And I don’t think I’d be alone,” says Redelinghuys, whose master’s studies were funded by UCT’s science faculty. “The real niche would be with retailers, because it would help them to understand their customers.

“If they could evolve to tell shoppers, ‘Based on what you bought last month, you should try this’, that would be powerful. And I would love that myself because I’m often standing in the wine section wondering, ‘Is it worth spending R200 to try something I may not like?’

“If a South African institution developed this, that’s where the real value lies.”