A $35 prix fixe menu was being served to customers at a restaurant called Trestle somewhere in the Jackson Square neighborhood of San Francisco. It’s likely that the patrons were unaware that they were also providing data to a machine learning model. A few years ago, a data science student called Tyler Knutson gathered tens of thousands of Yelp reviews from Bay Area eateries and fed them into a prediction model.
His wager was simple and a touch audacious: could the talk of amateur food critics on Yelp possibly foretell the decisions of Michelin’s famously secretive inspectors? The solution, it turned out, was more interesting than anyone expected.
Knutson’s method wasn’t fancy, but it was clever. He found that Yelp’s elite reviewers a self selected, somewhat mysterious group designated by Yelp’s own internal council behaved differently around restaurants that were going to receive or lose Michelin stars.
| Topic | Machine Learning & Michelin Star Prediction |
|---|---|
| Field | Data Science / Food & Gastronomy |
| Key Techniques | LGBMClassifier, Sentiment Analysis, NLP, Netnography |
| Notable Dataset | Kaggle Michelin Starred Restaurants (2018–2019) |
| Key Researchers Referenced | Tyler Knutson (NYC Data Science Academy), Christine Song (UBC CPSC330) |
| Platforms Used | Yelp, TripAdvisor, Kaggle |
| Michelin Guide Founded | 1900 (France) |
| Current Michelin 3-Star Restaurants Worldwide | Fewer than 150 |
| Reference | Michelin Guide Official |
Restaurants on the edge of receiving a star saw elite review volume jump by an average of 132% in the year before Michelin’s October announcement. Restaurants about to lose their stars saw elite reviews actually fall. The link was statistically significant, with a p value about 0.006. That’s not nothing. You sit up in your chair when you see a number like that.
In the meantime, a different group of University of British Columbia students attempted an entirely different strategy. Using a Kaggle dataset of 521 Michelin starred restaurants, Christine Song and her associates developed a classification model using a gradient boosting algorithm LGBM Classifier, for those keeping score. Their program predicted whether a restaurant should receive one, two, or three stars based on price level, cuisine type, and location coordinates.
When you take into account what they were working with, their 67% precision score on validation data may seem low. Of those 521 establishments, 419 had only one star, making the statistic glaringly unbalanced. The model hardly had enough instances to learn from because two and three star enterprises were so uncommon. A common overfitting issue that plagues small datasets is the possibility that the model was only memorizing patterns in the training data rather than identifying true signals.
Both projects have a deeper constraint that is worth pondering. The meal itself, which is likely the most important factor, was not available to either model. The braised short rib was not tasted by any algorithm. Whether the sommelier hesitated before suggesting a wine pairing was not detected by any neural network.
Instead of evaluating the sensory experience that Michelin claims to, machine learning, at least in these studies, was working with proxies, such as review sentiment, pricing brackets, and geographic coordinates. It’s similar to attempting to forecast the Academy Awards by examining the colors of movie posters. You might notice a pattern, but you should be cautious.
Nevertheless, the nexus between data science and gastronomy is quite fascinating. Latent Dirichlet Allocation and sentiment analysis were recently used by Turkish researchers to examine more than 21,000 Trip Advisor reviews of Turkish restaurants that are listed on Michelin.
Five recurrent themes emerged from their analysis of diners’ descriptions of luxury dining experiences ambience, overall experience, quality of food and service, local culinary identity, and what they termed complementary elements.
A framework based on psychologist Paul Ekman’s description of fundamental emotions was used to capture the emotional texture of those reviews, revealing patterns that could not be compiled by a single inspection. Thousands of subjective experiences are transformed into something that resembles a map of collective taste through this type of analysis.
Then there’s the question that no one likes to address aloud would Michelin want to be anticipated at all? The guide’s unpredictable nature and sense of heavenly judgment descending upon a kitchen contribute to its power.
The entire enterprise would lose part of its magic if a machine learning model could accurately predict the stars of the upcoming year. Gaming skilled chefs may begin to optimize on algorithmic signals instead of culinary sensibilities. Yelp reviews used to manipulate Google’s search results.
As of right moment, the models are still flawed; they are intriguing but not trustworthy enough to wager on a restaurant’s reputation. The UBC team admitted that by polluting their test set, they might have unintentionally broken the golden rule of machine learning. Although statistically sound, Knutson’s Yelp based forecasts only covered a small geographic area. These are not deployment ready tools; rather, they are research trials.
However, it’s difficult to ignore the sense that something is changing while observing this area. With an annual AI return on investment of more than 50 million euros, the Michelin Group already operates more than 200 AI use cases throughout its manufacturing activities.
It seems fairly certain that someone, either inside or outside of Michelin, will eventually train a model complex enough to make the anonymous inspectors a little uneasy if the parent firm is thus deeply involved in machine learning for tire inspection and supply chain predictions. No algorithm can determine whether that is beneficial for the fine dining industry.