Eco-score Prediction Deep Learning

We envisioned Green IA with two friends, with the goal of helping consumers make informed choices about food products🌱. You may already know this, but in addition to the Nutri-Score, an Eco-Score has been created, which is still experimental today but will gradually be added to the packaging of our food products 🛒. Like the Nutri-Score, the Eco-Score assigns a rating to a product ranging from A (very good) to E (very bad), based on its environmental impact 🌍: ozone layer depletion, impact on biodiversity, water consumption, etc. 💧

Our goal has therefore been to provide users with a prediction of this rating for all products in the OpenFoodFacts database, which is an open-access database, notably used by the Yuka app 📱.

We have thus developed a web application available on both computers and mobile devices 💻📱, allowing users to scan an item's barcode and then display the prediction result 🔍.

These predictions are made by a custom-built hybrid model, consisting of two multilayer perceptrons: one for numerical data, and another to process the outputs of the first perceptron along with the Transformers Bert model that handles our textual data 📊📚. In the output, we get the item's rating, which is accurate in 94% of cases ✅. In the remaining 6%, the error is always within one category, for example, we predicted a B instead of an A 🅱️➡️🅰️.

This model is included in an automatic pipeline that downloads data from the database every week, processes it, and then trains our model 🔄. If this model offers better performance than the previous one, it replaces it to make new predictions on all the new data 📈.