Recipe2Cuisine
Task
This is a data challenge I completed in 4 hours at Insight. The data is a small training set of about 39,000 recipes with labeled cuisines. The task is to design and execute a method to predict the cuisine of a recipe given only its ingredients. Also, I need to find the driving ingredients to characterize each major cuisine, so that I can write a guideline for an outsourced team to hand label the remaining unlabled recipes.
EDA
The data has 39774 non-null rows of recipes, 2 columns: cuisine and ingredients.
| id | cuisine | ingredients | |
|---|---|---|---|
| 0 | 10259 | greek | [romaine lettuce, black olives, grape tomatoes... |
| 1 | 25639 | southern_us | [plain flour, ground pepper, salt, tomatoes, g... |
| 2 | 20130 | filipino | [eggs, pepper, salt, mayonaise, cooking oil, g... |
| 3 | 22213 | indian | [water, vegetable oil, wheat, salt] |
| 4 | 13162 | indian | [black pepper, shallots, cornflour, cayenne pe... |
There are totally 20 classes of cuisines, and Italian food has the most recipes. The distributin of cuisine types is
Solution
To find the driving ingredients for each cuisine, I tried two methods: word frequency and impotant features.
Word Frequency
Basically, the more often the ingredients are used, the more important they are to the cuisine. I used nltk to find the most frequently used ingredients for each cuisine. For example, the driving ingredients for Italian food are:
However, it is difficult to identify the commonly used ingredients in all cuisines, such as salt, water, pepper, etc. This is hard to outsourced team to hand label the recipes.
Important Features
Generally, importance provides a score that indicates how useful or valuable each feature was in the construction of the boosted decision trees within the model. The more an attribute is used to make key decisions with decision trees, the higher its relative importance. The important features can be considered as the driving ingredients for the cuisines. To find the driving ingredient for each cuisine, I make the cuisine labels binary by setting one desired cuisine type as 1 and other 0, then I built tf-idf matrix for the ingredients as features and a random forest classification model to predict the binary outcome. Datasets were split 80% and 20% into train and test datasets respectively. Hyperparameter tuning was performed using Bayesian optimizaiton with 5-fold cross validation to obtain the optimized parameters for the model. Finanlly I exported the top 10 important features for the classification model, which can be considered as the most driving and representative ingredients for the desired cuisine:
After learning the top 10 driving ingredients for each cuisine from the classification results, the outsourced team would be able to label the rest of the recipes by identifying the driving ingredients from the guideline. The table below lists the top 5 driving ingredients for selected cuisines. Please see the complete list in the result file.
| Mexican | Southern_US | Indian | Chinese | Moroccan |
|---|---|---|---|---|
| corn tortillas | buttermilk | garam masala | soy sauce | ground cumin |
| flour tortillas | butter | ground turmeric | corn starch | ground cinnamon |
| salsa | chopped pecans | cumin seed | sesame oil | chickpeas |
| avocado | baking powder | curry powder | hoisin sauce | olive oil |
| jalapeno chilies | all-purpose flour | tumeric | oyster sauce | ground ginger |
| black beans | baking soda | ghee | light soy sauce | cinnamon |
| chili powder | vanilla extract | green chilies | dark soy sauce | paprika |
| sour cream | yellow corn meal | basmati rice | scallions | ground coriander |
| ground cumin | milk | mustard seeds | low sodium soy sauce | saffron threads |
| chopped cilantro fresh | salt | plain yogurt | fresh ginger | carrots |