Recipe-Reviews-Analysis

Project 3 for DSC 80 at UCSD

View the Project on GitHub jpuray/Recipe-Reviews-Analysis

The 40 Minute Rule!

In this project we studied the relationship between the minutes and average rating of a recipe. We wanted to learn if the time it takes for a recipe to be cooked had any effect on the average rating of that recipe. More specifically, we wanted to know:

Do shorter recipes (less than or equal to 40 mins) tend to have higher average ratings than longer recipes (40 mins or more)?

We believe answering this question will be useful to the creators who post there recipes online as this analysis would give insight on to what is the best cooking time for a dish, in order to get the best ratings.

Introducing the Data

The dataset we will be using for our analysis is made up of two datasets. RAW_RECIPES.csv which contains over one with over 70,000 recipes RAW_INTERACTIONS.csv which contains over 200,000 reviews of the recipes. Both datasets were obtained through the website food.com.

The variables we will be most intersted in particular are minutes in recipes and rating.

Breakdown of datasets:

RAW_RECIPES.CSV

Variables Description
name Recipe name
id Recipe Id
minutes Minutes to prepare recipe
contributor_id Contributor Id
submitted Date Submitted
tags List of tags
nutrition Nutrtion facts
n_steps Number of steps
steps List of steps
description Description of Recipe

RAW_INTERACTIONS.CSV

Variables Description
user_id User Id
recipe_id Recipe Id
date Date review was written
rating Rating of recipe
review Review

Cleaning and EDA

Data Cleaning

To start by loading in the data and merging the two datasets.


# loading data 
recipes = pd.read_csv('data/RAW_recipes.csv')
interactions = pd.read_csv('data/RAW_interactions.csv')
# merging DataFrames together 
merged = recipes.merge(interactions, left_on='id', right_on='recipe_id', how='left')

The resulting dataframe has over 200,000 rows with 17 columns. This makes sense as we are merging the interactions dataset to the recipes dataset. The recipes will be duplicated but the interactions data are unique which we will use to calcualte averge rating later on.

We noticed that the merged dataset, had many values in the rating column with the value 0.

  rating
5 169676
4 37307
0 15035
3 7172
1 2870
2 2368

This could pose a problem for our analysis since we will be using the average rating of recipes and values of 0 would decrease our averages. After looking at the website where the data had been scraped from (food.com) we realised that it was not possible to rate a recipe less than 1 star. Upon further analysis we identified a section labeled ‘tweaks’ which contained reviews but no ratings. Thus, we concluded that the rows with 0 star rating were from the ‘tweak’ section and that those stars are acutally supposed to be missing by design. Therefore, we changed the ratings with 0 to NaN so that these rows did not affect our average ratings.

We then grouped all the recipes with the same id and calculated the mean to get the average rating and added it as a column labled avg_rating.

avg_rating = merged.groupby('id')['rating'].mean()
merged['avg_rating'] = merged['id'].apply(lambda x: avg_rating[x])

We then added the column time label which contains the string values ‘40<=’ and >40’, then dropped the following columns as they would not be used for further analysis:

  1. ‘submitted’,
  2. ‘user_id’,
  3. ‘recipe_id’,
  4. ‘date’,
  5. ‘rating’ (already used to calculate average rating)

Removing outliers

When looking at the distributions of our variables we noticed the variable minutes some outliers that we wanted to deal with before our analysis. Plotting the distribution of the minutes column gave us the follwing histogram:

In order to identify outliers we used the Inter quartile range (IQR) method. In which we define outliers as any value less than then the (25th percentile * 1.5) or greater than (75th percentile * 1.5). Then when a value was detected as an outlier it was removed from the dataset as these values would be unrepresentative of our general population of recipes. After removing the outliers our new histogram became:

Afterwards we are left with the following DataFrame of shape (75791, 13):

  name id minutes contributor_id tags nutrition n_steps steps description ingredients n_ingredients avg_rating time label
0 1 brownies in the world best ever 333281 40 985201 [‘60-minutes-or-less’, ‘time-to-make’, ‘course… [138.4, 10.0, 50.0, 3.0, 3.0, 19.0, 6.0] 10 [‘heat the oven to 350f and arrange the rack i… these are the most; chocolatey, moist, rich, d… [‘bittersweet chocolate’, ‘unsalted butter’, ‘… 9 4 <=40
1 1 in canada chocolate chip cookies 453467 45 1848091 [‘60-minutes-or-less’, ‘time-to-make’, ‘cuisin… [595.1, 46.0, 211.0, 22.0, 13.0, 51.0, 26.0] 12 [‘pre-heat oven the 350 degrees f’, ‘in a mixi… this is the recipe that we use at my school ca… [‘white sugar’, ‘brown sugar’, ‘salt’, ‘margar… 11 5 >40
2 412 broccoli casserole 306168 40 50969 [‘60-minutes-or-less’, ‘time-to-make’, ‘course… [194.8, 20.0, 6.0, 32.0, 22.0, 36.0, 3.0] 6 [‘preheat oven to 350 degrees’, ‘spray a 2 qua… since there are already 411 recipes for brocco… [‘frozen broccoli cuts’, ‘cream of chicken sou… 9 5 <=40
6 millionaire pound cake 286009 120 461724 [‘time-to-make’, ‘course’, ‘cuisine’, ‘prepara… [878.3, 63.0, 326.0, 13.0, 20.0, 123.0, 39.0] 7 [‘freheat the oven to 300 degrees’, ‘grease a… why a millionaire pound cake? because it’s su… [‘butter’, ‘sugar’, ‘eggs’, ‘all-purpose flour… 7 5 >40
7 2000 meatloaf 475785 90 2202916 [‘time-to-make’, ‘course’, ‘main-ingredient’, … [267.0, 30.0, 12.0, 12.0, 29.0, 48.0, 2.0] 17 [‘pan fry bacon , and set aside on a paper tow… ready, set, cook! special edition contest entr… [‘meatloaf mixture’, ‘unsmoked bacon’, ‘goat c… 13 5 >40

Univariate Analysis

First we will plot the distribution of the variables minutes and avg_rating in order to get an idea of what the distriubtions look like.

Plotting the distribution of the minutes gives us the following histogram:

Plotting the distribution of the avg_rating gives us the following histogram:

The average ratings in the dataset are skewed left and tend to have more positive ratings as opposed to negative ones. Over 60% of the recipes have an average rating of 5 while less than 5% have a rating of 3 or less.

Bivariate Analysis

Next we will plot the conditional distributions of the variables on whether or not the recipe is less than 40 minuetes or less or not.

Looking at the two distributions there seems to be a disproportionate amount of recipes on the site that take less than 40 minutes to cook, but there relative shapes look similar.

Interesting Aggregates

Next we will create a pivot table in order to compare the average rating of the shorter recipes vs the longer recipes over the number of ingredients to get a better idea of how the averages may differ.

time label 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
<=40 4.792 4.74624 4.71655 4.70764 4.71222 4.69611 4.67469 4.66409 4.68101 4.66748 4.67962 4.69613 4.71457 4.66231 4.68861 4.68911 4.72753 4.72819 4.71846 4.77003 4.72861 4.82624 4.83673 4.69231 4.11111 4.89853 4.8 5 4.90476 5 nan nan 5
>40 5 4.63603 4.73703 4.65007 4.71809 4.64523 4.65759 4.64848 4.63874 4.65828 4.69217 4.6648 4.69096 4.65559 4.68107 4.66201 4.64584 4.75513 4.67791 4.67346 4.68863 4.89439 4.82418 4.79658 4.80882 4.81481 4.59048 4.88571 4.94118 4.69565 5 5 nan

Calculating the differnece in means across ingredients:

pivot.diff().mean(axis=1)

We get that the difference in means is -0.006732 showing a slightly higher average rating for shorter cooking time recipes.

Assessment of Missingness

Our dataset, Recipes and Ratings, contained three columns in which there were missing values. We can determine these columns using pandas functions:

  0
user_id 0
recipe_id 0
date 0
rating 0
review 169
  0
name 1
id 0
minutes 0
contributor_id 0
submitted 0
tags 0
nutrition 0
n_steps 0
steps 0
description 70
ingredients 0
n_ingredients 0

We can see that there are missing values in the name, description, and review columns of the interactions and recipes dataframes.

  1. Since there is only one missing value in the name column, we cannot make a solid determination regarding the type of missingness of that column. In order to make a concrete determination of the type of missingness, we would likely need to scrape more recipe submissions and check which other submissions also had missing name values.  
  2. For the review column, we determined that the missingness type is likely MAR. That is, the missingness of the review column is dependent upon another column. Likely columns for which the missingness of reviews would be dependent on includes the rating column. Unless the user who left a rating on the recipe felt especially strong about the recipe (e.g. strongly disliked (1) or strongly liked (5)), then they would be unlikely to go into written detail about their thoughts on the recipe.  
  3. In terms of the description column, there is a possibility of the description being NMAR (Not Missing at Random). Some recipe submissions simply do not require descriptions for the recipe as determined at the discretion of the user who submitted it. Thus, the missingness of the description is independent of any of the other columns of the recipes dataframe.

After cleaning, merging, and adding necessary variables to our dataframe, our dataframe appeared as so:

  name id minutes contributor_id tags nutrition n_steps steps description ingredients n_ingredients avg_rating time label
0 1 brownies in the world best ever 333281 40 985201 [‘60-minutes-or-less’, ‘time-to-make’, ‘course… [138.4, 10.0, 50.0, 3.0, 3.0, 19.0, 6.0] 10 [‘heat the oven to 350f and arrange the rack i… these are the most; chocolatey, moist, rich, d… [‘bittersweet chocolate’, ‘unsalted butter’, ‘… 9 4 <=40
1 1 in canada chocolate chip cookies 453467 45 1848091 [‘60-minutes-or-less’, ‘time-to-make’, ‘cuisin… [595.1, 46.0, 211.0, 22.0, 13.0, 51.0, 26.0] 12 [‘pre-heat oven the 350 degrees f’, ‘in a mixi… this is the recipe that we use at my school ca… [‘white sugar’, ‘brown sugar’, ‘salt’, ‘margar… 11 5 >40
2 412 broccoli casserole 306168 40 50969 [‘60-minutes-or-less’, ‘time-to-make’, ‘course… [194.8, 20.0, 6.0, 32.0, 22.0, 36.0, 3.0] 6 [‘preheat oven to 350 degrees’, ‘spray a 2 qua… since there are already 411 recipes for brocco… [‘frozen broccoli cuts’, ‘cream of chicken sou… 9 5 <=40
6 millionaire pound cake 286009 120 461724 [‘time-to-make’, ‘course’, ‘cuisine’, ‘prepara… [878.3, 63.0, 326.0, 13.0, 20.0, 123.0, 39.0] 7 [‘freheat the oven to 300 degrees’, ‘grease a… why a millionaire pound cake? because it’s su… [‘butter’, ‘sugar’, ‘eggs’, ‘all-purpose flour… 7 5 >40
7 2000 meatloaf 475785 90 2202916 [‘time-to-make’, ‘course’, ‘main-ingredient’, … [267.0, 30.0, 12.0, 12.0, 29.0, 48.0, 2.0] 17 [‘pan fry bacon , and set aside on a paper tow… ready, set, cook! special edition contest entr… [‘meatloaf mixture’, ‘unsmoked bacon’, ‘goat c… 13 5 >40

Since we utilize the avg_rating column in our data analysis, and there are missing values (as shown below), we determined it necessary to test the missingness dependency of avg_rating on the column, minutes.

To test if the avg_rating column is MAR, we execute a permutation test. The code to do so is included below:

def test_missing(col,k=1000):
    df = cleaned_food.copy() # create a copy of the cleaned dataframe
    
    # calculating the observed test statistic for our permutation test
    non_missing_mean = df[~df['avg_rating'].isna()][col].mean()
    missing_mean = df[df['avg_rating'].isna()][col].mean()
    obs_diff_mean = abs(non_missing_mean-missing_mean)
    
    results = [] #empty list to store results
    
    #perform k iterations of the permutation test 
    for i in range(k):
        df[col] = np.random.permutation(df[col]) # shuffle the specified column
        non_missing_mean_test = df[~df['avg_rating'].isna()][col].mean() # non-missing rows 
        missing_mean_test = df[df['avg_rating'].isna()][col].mean() # missing rows 
        diff_mean = abs(non_missing_mean_test-missing_mean_test) # compute the difference in means
        results.append(diff_mean) # append test statistic to results list
    p_val = (np.array(results)>=obs_diff_mean).mean() # compute the p value based on the results
    return p_val

When we run the function above with the input column specified as ‘minutes’, we get the following output:

test_missing('minutes')
0.0

That is, we conclude that at the significance level of p = 5%, that there is likely a dependency on the ‘minutes’ column.

We then test the dependency of the avg_rating column on another column, ‘n_ingredients’, which includes the number of ingredients that the recipe calls for. We do so in a similar manner to our test for minutes:

test_missing('n_ingredients')
0.06

In this case, we conclude that it is unlikely that the avg_rating column depends on ‘n_ingredients’ because it is above the 5% significance level.

Hypothesis Testing

Recall, we are looking into the question:

What is the relationship between the cooking time and average rating of recipes?

To complete an in-depth analysis of our question based on the data, we chose to run a permutation test. Details of our test include:

Null Hypothesis: Ratings for recipes that take less than or equal to 40 minutes to complete have the same ratings as recipes that take greater than 40 minutes to complete.

Alternative Hypothesis: Recipes that take less than or equal to 40 minutes to complete tend to have higher ratings

We chose these as our null and alternative hypotheses because we found that in plotting the relationship between minutes and average rating, most of the high ratings for recipes could be found in the recipes that took less than 40 minutes. Thus, we decided to compare the two groups (>40 and <=40) in our test.

Test Statistic: Difference in Means

Because our Alternative Hypothesis specifies that ratings are higher for recipes with durations less than or equal to 40 minutes, we choose an unsigned difference in means to compare the two.

P-Value: 5% significance level

We chose the conventional significance level for our p-value.

For ease of permutation testing, we added a column to our dataframe that specifies whether the recipe falls within the <=40 or >40 group.

To perform our permutation test, we created a function that takes the number of iterations of the test we’d like to perform (with a default value of 100), and returns the p-value for our test. Code is as follows:

def perm_test(k=100):
    df = cleaned_food.copy() # create a copy of the cleaned_food dataframe
    rows_greater_40 = df[df['time label']== '>40'] # isolate >40 minute duration values in a dataframe
    rows_less_40 = df[df['time label']=='<=40'] # isolate <=40 duration values in a dataframe
    
    obs_stat = rows_less_40['avg_rating'].mean() - rows_greater_40['avg_rating'].mean() #calculate observed difference in means
    
    results = [] # create empty list for results
    
    #perform k iterations of test
    for i in range(k):
        df['time label'] = np.random.permutation(df['time label']) # shuffle time label column
        row_g40 = df[df['time label']=='>40'] # get dataframe of >40 minute recipes
        row_l40 = df[df['time label']=='<=40'] #get dataframe of <= 40 minute recipes
        
        ts = row_l40['avg_rating'].mean()- row_g40['avg_rating'].mean() #calculate the current test statistic
        
        results.append(ts) #append test statistic to results list
    p_val = (np.array(results)>=obs_stat).mean() #calculate p value for our test
    return p_val

When we run this function with the default values, we get the following output:

perm_test()
0.0

Because our p-value for the test is below our pre-determined significance level of 5%, we reject the null hypothesis. That is, we conclude that it is more than likely that the two groups (duration > 40 and duration <= 40) result in different average ratings.