Uptaker, predicting the most popular recipes on a menu by Ethan Potter, Hai Nguyen & Irene Iriarte Carretero
In Gousto, our data products are used in a wide range of business use cases such as understanding customer tastes for recipe recommendations, optimising our factory picking lines or providing accurate forecasts to minimise food waste. In this presentation, we will walk you through two examples of Gousto data products that have been developed separately but work very well in harmony:
Rouxcommender, our recipe recommendation engine, and RPD, our recipe popularity forecaster. Rouxcommender is a family of Gousto’s recommender systems that aims to bring the best of our recipes to the customers. To achieve such an ambitious goal, Rouxcommender continuously learns about our customers’ unique tastes from consumption behaviours with both implicit and explicit data, while also building suitable representations of our recipes and ingredients (embeddings) to tackle the infamous cold-start problem. In this talk we would like to tell more about Rouxcommender as well as how we tackle some of the interesting challenges in personalisation at Gousto.
While Rouxcommender optimises the customer experience, as “a data company that loves food” we also need to optimise our supply chain. Given the short shelf life of food produce, forecasting the popularity of a recipe or ingredient accurately is paramount for any food-based supply chain. It directly determines the number of ingredients to buy, meaning, if forecasted inaccurately, it can result in food waste, the inability to fulfil demand, and increased costs. We show that with recipe embeddings, we can accurately predict the popularity of a recipe we haven’t sold before, on a menu that hasn’t existed before.
Supported by
