How does Nutrition influences Food ratings?
Introduction
Food.com is a website which allows chefs share their tasty recipes with the world and internet users globally can give their comments on the recipe and a rating from 1-5 to give a unbiased view on the published recipes and answer whether these recipes are worth cooking up or not. Food.com has published a publically accessible database which has metrics like ingredients, nutritional facts, a description of the recipe, and much more. My focus with this project is to predict what a user will rate a particular recipe based on the recipes nutritional data. I chose nutritional facts to predict the ratings because I was curious what type of foods people like making for example do they like making foods like brownies ro cakes which are generally unhealthy, or foods that are high in protein and low in sugar as a healthy option.
by Ehsan Kabeer (eskabeer@umich.edu)
Data Cleaning and Exploratory Data Analysis
Cleaning
Before I got into any predictive model bulidng I had to clean up the dataset. The dataset was initially was split into two tables, one with the actual recipes and their respective details and another with the users ratings and their thoughts on the recipes. What I did was merge the two tables with the left merge, using the recipe ids. After this I dropped the null rows from table and filled the ratings in the table with a value of 0 as null (since scale is 1-5).
After I did this I wanted to get rid of columns from the table that weren’t relevant for the model building process. This resulted in the removal of columns: recipe_id, user_id, date, submitted, minutes, tags, n_steps, steps, description, ingredients, review, contributor_id, and n_ingredients.
After this I needed to split the nutrition facts into seperate columns so I can use them. Initially they were in a list format; this included information like calories (#), total fat (PDV), sugar (PDV), sodium (PDV), protein (PDV), saturated fat (PDV), carbohydrates (PDV)] (PDV is percentage of daily value). I split them each into individual columns using regex and after that I dropped the nutrition column since I didn’t need it.
For visualization purposes I made new columns for each nutrition fact that split it into 3 bins, for example for the Protein(PDV) I split into low protein, moderate protein, and high protein with low being less than 5%, moderate being less than 19% and high being above that; this will allow the visualization I am going to be showing to be more clean.
| name | id | rating | calories | total_fat(PDV) | sugar(PDV) | sodium(PDV) | protein(PDV) | saturated_fat(PDV) | carbohydrates(PDV) | rating_avg | fatClass | satFatClass | sugarClass | sodiumClass | proteinClass | carbClass | calClass |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 brownies in the world best ever | 333281 | 4 | 138.4 | 10 | 50 | 3 | 3 | 19 | 6 | 4 | Moderate_fat | Unhealthy | High_sugar | low_sodium | low_protein | Moderate_carb | Moderate_cal |
| 1 in canada chocolate chip cookies | 453467 | 5 | 595.1 | 46 | 211 | 22 | 13 | 51 | 26 | 5 | High_fat | Unhealthy | High_sugar | high_sodium | Moderate_protein | High_carb | High_cal |
| 412 broccoli casserole | 306168 | 5 | 194.8 | 20 | 6 | 32 | 22 | 36 | 3 | 5 | High_fat | Unhealthy | low_sugar | high_sodium | High_protein | Low_carb | Moderate_cal |
| 412 broccoli casserole | 306168 | 5 | 194.8 | 20 | 6 | 32 | 22 | 36 | 3 | 5 | High_fat | Unhealthy | low_sugar | high_sodium | High_protein | Low_carb | Moderate_cal |
| 412 broccoli casserole | 306168 | 5 | 194.8 | 20 | 6 | 32 | 22 | 36 | 3 | 5 | High_fat | Unhealthy | low_sugar | high_sodium | High_protein | Low_carb | Moderate_cal |
Univariate Analysis
With this graph we are exploring the distribution of fat level in a recipe among the differnt ratings, the fat level was split into low fat, moderate fat, and high fat. With this visualization I wanted to see if consumers prefer a more low fat option for the recipes they are cooking up as opposed to a high fat ones. From the graph you can see that five star rated items tend to be on the lower fat side, and if you look closely at the 3 star distribution, high fat items slightly are above in percentage compared to lower fat options. So from this visualization there is slight trend pointing to the fact that low fat items are the preferred recipes.
With this graph we are exploring the distritbution of calorie level in a recipe among the different ratings, just like with fat this was split into low calorie, moderate calorie, and high calorie. With this graph I wanted to further investigate whether consumers prefer healthier or less healthy options. This graph again shows that the five star items are teh lower claorie items with low calorie being highest percentage and moderate being a close second. The lower ratings like four and three star both had high calorie as top, which further soldifies that trend that healthier options tend to be more highly rated.
Bivariate Analysis
For this graph I took a simple random sample from data of 20,000 so the graph is more easily interpretable. For this graph I am how the sodium(PDV) infuences the average ratings of the recipes. I wanted to see if consumers preferred salty snacks when considering what to make. From the graph we can see that as salt content gest higher rating seem to get higher, which I was suprised by since in the other graphs it was seen that consumers preferred more healthy recipes.
Pivot table
satFatClass represents the classification of the saturated fat content for a particular recipe. With the categories being unhealthy, caution, and heart-healthy. These designations were made since saturated fats are the actual unhealthy fats in food.
| satFatClass | rating_avg | calories |
|---|---|---|
| Caution | 1 | 246.852 |
| Caution | 2 | 284.562 |
| Caution | 3 | 244.674 |
| Caution | 4 | 262.265 |
| Caution | 5 | 238.242 |
| Heart_healthy | 1 | 160.828 |
| Heart_healthy | 2 | 185.624 |
| Heart_healthy | 3 | 196.456 |
| Heart_healthy | 4 | 181.42 |
| Heart_healthy | 5 | 173.945 |
| Unhealthy | 1 | 660.587 |
| Unhealthy | 2 | 619.091 |
| Unhealthy | 3 | 620.657 |
| Unhealthy | 4 | 574.03 |
| Unhealthy | 5 | 572.7 |
This grouped table is grouped on the saturated fat class and average rating, and we are getting the mean calories for each group. What I wanted to explore here is to see the behavior of consumers on how much calories they prefer to be in a recipe based on how healthy that recipe is. A interesting trend that can be gathered from this table is that when consumers choose to make a unhealthy or even moderately unhealthy recipe they prefer the lower calorie options, but when they choose to make a healthy food, they were fine with the recipe being higher in calories.
Imputation
One of the columns that had to be imputed was the rating column. Some of the rows in the rating columns had a value of 0 which is a invalid rating so what I chose to do is I made those rows null for their rating, as mentioned before then I did a mean imputation for the null ratings by their respective recipes, which would allow me to keep the general distribution of the dataset while not losing any data from table (some recipes only had 0 ratings, so they were dropped completely from data after mean imputation). In the figures below we show the distribution of ratings before the imputation and after.
Framing a Prediction Problem
Prediction Problem
We are looking to predict the rating of a recipe based on the nutritional facts of a recipe namely calories (#), total fat (PDV), sugar (PDV), sodium (PDV), protein (PDV), saturated fat (PDV), carbohydrates (PDV). So, given all the nutritional data of a particular recipe, can we accurately predict its rating? We will be using a Linear Regression model to predict the rating using nutritional data since the rating is a continuous variable.
Response variable
Since the rating of a recipe indicates how good that recipe is for others cook and eat, we will use it as our response variable to explore how nutritional metrics of a recipe effect its rating.
Predictor variables
The predictor varibles I chose were the following: calories (#), total fat (PDV), sugar (PDV), sodium (PDV), protein (PDV), saturated fat (PDV), carbohydrates (PDV). Since these values correlate to how healthy or unhealthy, or how filling or non-filling a recipe is, they could influence the rating of a recipe greatly. Since these values are derived from the ingredients of a recipe it is known before the recipe is rated.
Metric to Evaluate Model
Since we are using a Linear regression model, mean sqaured error is definetly the appropriate metric to evaluate model performace, as it is the standard for Linear Regression and it gives a direct way of interpreting performance of model.
Baseline Model
Model
Linear Regression was the model selected for this problem since its easy to interpret based on its coefficients.
Features in Model
The features used in the model were calories (#), total fat (PDV), sugar (PDV), sodium (PDV), protein (PDV), saturated fat (PDV), and carbohydrates (PDV) which are all quantitative so a total of 7 quantitative variables. No ordinal or nomial features were used so there was no need for any encoding for this model.
Performance
The performance of this model was measured using mean squared error as mentioned before. It takes a sum of teh magnitude between predicted and actual values in test set, with a lower MSE meaning better performance. With the base model a MSE of 0.255947 was acheived. Since the scale of values for the ratings is 1-5 this performance is relatively good but there is still room for improvement which we will try and do below.
Final Model
Before I moved on to actual improvements on model, I used Standard scalar on model since some of the features, namely calories, are on a different and larger scale than the other features, so standard scaler will bring these to a mean of 0 and a standard deviation of 1 which will allow easier interpretation of what the important features in the model. In terms of the steps I took to actually improve the model, I used polynomial features to handle potential cases of non-linear relationships showing up in the data and I also used ridge and lasso regression so handle unimportant features. By spotting out non-linear relations with the model and also eliminating unimportant features the model will be able to capture the more unexpected relations between nutrition data and ratings, which will produce better results.
After trial and error and cross-validation Ridge came out on top as teh better modeling algorithm, with lasso actually performing worse than base model. The hyperparamters that performed best were I chose degree as one of the hyperparameters since I wanted to see which degree will capture the non-linear relations apprent in the data. I also used Ridge alpha as hyperparamter to find the perfect penalization of the coefficients.
The MSE of the final model is 0.25653 which is a slight improvement over 0.25594. Even though this is a small improvement its still some improvement which is great, and shows that final model does better on unseen data than the base model