Food Categorization

Let's start looking at the dataset and how we decide to analyse it.

The data set we analysed comes from the Open Food Facts database that gathers information and data on food products from around the world and it gets constantly updated. The data set includes almost 2 GB of data, giving insights on nutritional values, additives content, brands and many other information about grocery products. Unfortunately, being the dataset curated on a voluntary basis, many data is missing.

In our analysis, we decide to consider the impact on health of the food products stored in the dataset, by first trying to assess which food category is more dangerous for people. We could come up with obvious results, but it is always better to check it! The food categories we considered are 9, and correspond to the ones defined by Programme national nutrition santé (PNNS): Beverages, Cereal and Potatoes, Composite Foods,Fat and Sauces, Fish and Meats and Eggs, Fruits and Vegetables, Milk and dairy products, Salty Snacks and Sugary snacks. In the table below there is an example of how we divided the different products according the above cited categories.

Nutritional Values For Category

Which nutritional value do we care ?

The goal of our project is to give the consumer advices about the healthiness of the products for the different categories according to a specific score. First, we explored the dataset, to find the intrinsic relationships that connect each category. From our analysis, many confirmations of what the common sense says have been found. For example, we show the average distributions of the categories for the variables carbohydrates and fats. What we can observe is that categories such as salty snacks, sugary snacks and potatoes have low fat content (generally), while high carbohydrate contents. Other categories, such as dairy products, have higher fat content than carbohydrates, and so on.

We decide to create our own score that includes three features available in the data set: Nutritional score, Energy over 100g and number of palm oil ingredients.
In order to fill the missing data for the nutritional score and the product categories, we decided to use a Random Forest Classifier based on the following features: energy, fat, sugars, proteins, carbohydrates and salt for 100 g of products.
We perfored a t-SNE dimensionality reduction to check whether the space in which we are representing the data is sufficiently good to distinguish between categories. It is possible to see that the categories are well separated one another: this indicates the possibilities to build a classifier that categorizes unlabeled data. With our Random Forest we managed to achieve a F1 score of 0.87. Moreover, we can observe that subclusters are present in the data: this could indicate that a finer representation of food categories is needed to give better health advices! This categorization could be improved by national authorities to make people live healthier lives.

Similarly, we constructed another Random Forest to predict the nutritional grade of unlabeled products. The nutritional grade is a French score that indicates how a product affects health, and it is one of the main elements of the our food advisor. Since it is missing in most of the non-French data, we extended it. The performance of the classifier was in this case 0.77 (F1 score). Most of the misclassified data was represented in adjacent score to the real one: hence the error made during the classification was generally low. This is our first achievement: we provide the consumers some missing values for the nutritional score that he can count on to choose what to eat. The next step is to expand this French nutritional grade, to get to our food advisor.

Product Advisor Based on the Category

Let's choose what you wanna eat

In the final part, we want to give you a product advisor. Imagine being at the supermarket, in doubt whether choosing pork ribs or chicken for dinner. It would be handy to have a score that tells you which product is more likely to have a negative impact on your health and on the environment.

The product score is based on the following metrics: up to 3.5 points are given by the French nutritional score, up to 1 point by the absence of palm oil in the ingredients of the product (that is environmentally dangerous) and another 1.5 points inverse proportionally to the energetic content of the product. You can have a taste of how our score works in the following table, in which you can view the 20 best products for each above-mentioned category.

Click here to visualize the table.

As you may have noticed, some products are misclassified, and some others have a strangely high score: this is either due to an error in the classification we did or to an error in the user insertion of the product. In any case, it is one of the limitations of this dataset!

Are you satified ? 😁

Food Category VS health

What is the distribution of the Nutri-Score within each food category ?

As expected... the healthier products are fruits and vegetables ! And obsiously you should avoid to eat snacks and beverages.

Interestingly, the sugary snacks are worse for health than salty snacks or the fatty products. Indeed the ratio of e products is 49% for sugary snacks and 14% for salty snacks or 22% for fat and sauces. Moreover, the sum of e + d ratios is also higher for sugary snacks.

This could demonstrate that the real enemy is maybe more sugar than salt or fat which contrasts with the standard nutritional recommendations which usually emphasize on fat.

Brands VS Health

Does Carrefour products healthier than Leclerc ones ?

                            Yes... Unfortunately, more than 50% of the products sold in France have a bad nutrition grade (d or e).           The healthy products tagged with the grade a only represent 13% of the ~100,000 products in the dataset.                          As seen previously, we observe that for almost all the biggest food brands present in France more than 40% of the products are "junk" food (products with labels d or e).           We don't eat that well in France...           

          The exception is Picard in which products tagged with d or e represent one third of the total number of products (which is already quite high!).           Actually, Picard mainly sells frozen food which could be expected to be healthier than the products sold by Netto, that is a "top budget" brand in France with cheap, mainly industrial. products           

          For each brand the score will be the average of the ratios for each nutri-score (a,b..) weighted by the score assigned to each nutri-score (5,4..).           A score of 5 for a brand means that 100% of the products sold by the brand are a (very healthy).           

          Below is displayed the 10 best (highest scores) and the 10 worst (lowest scores) brands, regarding nutrition benefit.                                  As we could expect, eating chocolate is not good... But, suprisingly the 6 worst brands are all chocolate ones except Monin which sells sirop.           The best brands are D'Aucy - which sells a lot of vegetables and fruits in cans - and Charles et Alice - which sells fruits in cup.                      

Palm oil in France

Does french products contain palm oil ?

                    A large majority (87%) of the products in France doesn't contain palm oil !
Ok... it can be addressed that a lot of products don't need oil along its fabrication process.                              For all the brands more than 70% of the products does NOT contain palm oil.           

          
• However, 27% of the products made by Leclerc contain (or may contain) palm oil. It is a high ratio!
          
• Picard and Monoprix are the more "environmental" brands regarding palm oil presence in their products.