# Cupcakes vs. Muffins

One thing I’ve learned from my PhD at Tufts is that I really enjoy working data wrangling, visualization, and statistics in R. I enjoy it so much, that lately I’ve been strongly considering a career in data science after graduation. As a way to showcase my data science skills, I’ve been working on a side project to use webscraping and multivariate statistics to answer the age old question: Are cupcakes really that different from muffins?

Honestly, I can’t even quite remember how this idea came to me, but it started in a discussion with Dr. Elizabeth Crone about why more ecologists don’t use a statistical technique called partial least squares regression. We wanted a fun multivariate data set that could illustrate the different conclusions you might get depending on the statistical method you use. Around the same time, I came across a blog post by @lariebyrd explaining machine learning using cake emoji. And that somehow led to me webscraping every single muffin and cupcake recipe on allrecipes.com.

I just finished the webscraping bit of the project this weekend. I’m not going to reproduce the code here, but rather address some of the challenges I faced, and some things I’ve learned so far. You can check out my R notebook and a .rds file of all the recipes over on github

# Getting started on webscraping

I followed this wonderful tutorial from José Roberto Ayala Solares to get going. Necessary tools include the tidyverse, the rvest package for webscraping, and a chrome plugin called SelectorGadget. Going through the example in the tutorial was really helpful for getting a hang of what is and isn’t easy/possible.

# Choosing a data source

I specifically chose allrecipes.com over, say geniuskitchen.com (an excellent recipe site) because of the way it categorizes and structures recipes. When I search “cupcake” on geniuskitchen.com, I get 1830 results (yay!), but a bunch of them are links to videos, articles, reviews, blog posts, food porn albums, and other things that are not recipes (boo!). Allrecipes.com, on the other hand, gives me the following URL: https://www.allrecipes.com/recipes/377/desserts/cakes/cupcakes/

This is ALL THE CUPCAKES. From there, I played around with SelectorGadget to make sure it was going to be easy to drill down to just the links to the actual recipes, and yes, it was.

I also checked that the ingredients list of each recipe was going to be easy to scrape. They all have the same format, and some brief testing convinced me that I’d be able to figure out how to pull out ingredients.

The takeaway is that for this project I had my choice of many websites, but I specifically picked one that would make my life easier because of its html structure.

# Don’t scrape too fast!

## (Sys.sleep() is your friend)

So once I figured out how to pull in all the links to all the cupcake recipes, I started scraping ingredients from a small sample of them. After a few debugging runs, allrecipes.com stopped responding and I just kept getting error messages. After pulling my hair out trying to figure out how I broke my code, I realized that my IP was being blocked! Because of the speed of accessing the website or how many links I accessed, my IP was suspected of being a bot or something and was temporarily (whew) blocked. I turned to twitter and was reccomended an easy fix—create a custom read_html_slow() function that included Sys.sleep(5) which just makes R wait 5 seconds in between reading websites.

library(rvest)
Sys.sleep(5) #wait 5 seconds before returning output
return(output)
}

# Create your own custom helper functions

A great side-effect of creating your own functions like read_html_slow() is making your code more readable. Instead of a for-loop that calls Sys.sleep(5) after every iteration of read_html(), I now have one function that does it all and can be easily used in conjunction with map() from the purrr package.

My read_html_slow() function would still occasionally encounter errors like when it would encounter a broken URL. When reading in a whole list of URLs, one broken URL would mess up the whole list. I ended up expanding on read_html_slow() to make read_html_safely() which would output an NA rather than throwing an error if a URL was broken.

library(purrr)
read_html_safely <- possibly(read_html_slow, NA) #from the purrr package

I also created a str_detect_any() which allows you to check if a string is matched by any of a vector of regular expressions. I show how I use this in the next section

library(stringi)
str_detect_any <- function(string, pattern){
map_lgl(string, ~stri_detect_regex(., pattern) %>% any(.))
#map_lgl is from purrr, stri_detect() is from stringi
}

# Work on random samples

There were something like 200 cupcake recipes and another 100 muffin recipes on allrecipes.com, which takes a long time to read in. Rather than working on the whole data set, I used sample() on my vector of recipe URLs to take a manageable sample of recipes to work on. After working through a few different random subsets, I reached a point where I was happy with how my code was working. Only then did I read in the entire data set.

# Webscraped data is messy

## (People make weird baked goods)

Once I had all my ingredients for muffins and cupcakes in a data frame, I needed to standardize the ingredients. For example “8 tablespoons butter, melted” and “1/2 cup unsalted, organic, non-GMO, gluten-free, single-origin butter” both needed to get converted to “1/2 cup butter.” This is where the combination of mutate(), case_when() and str_detect() really came in handy to make readable, debuggable code.

mutate() is a function from the dplyr package (part of tidyverse) for adding new columns to data frames based on information in other columns. Here, I used it to take the ingredient descriptions and turn them into short, concise ingredients. str_detect() is from the stringr package and takes a string and a regular expression pattern and outputs TRUE or FALSE. Finally, case_when() is also from dplyr and provides a readable alternative to insane nested ifelse() statements. For example:

library(tidyverse)
df <- tibble(description = c("1/2 cup unsalted, organic, non-GMO, gluten-free, single-origin butter",
"1 cup buttermilk",
"4 cups sugar",
"4 cups slivered almonds",
"1/2 cup chopped walnuts",
"1 teaspoon salt",
"25 blueberries"))

#all nuts should match one of these patterns
nuts <- c("almond", "\\w*nut", "pecan")

df %>%
mutate(ingredient = case_when(str_detect(.$description, "butter") ~ "butter", str_detect(.$description, "milk")   ~  "milk",
str_detect(.$description, "sugar") ~ "sugar", str_detect(.$description, "salt")   ~  "salt",
str_detect_any(.\$description, nuts) ~  "nut",
TRUE                                ~  as.character(NA)
)
)
## # A tibble: 7 x 2
##   description                                                         ingredient
##   <chr>                                                               <chr>
## 1 1/2 cup unsalted, organic, non-GMO, gluten-free, single-origin but… butter
## 2 1 cup buttermilk                                                    butter
## 3 4 cups sugar                                                        sugar
## 4 4 cups slivered almonds                                             nut
## 5 1/2 cup chopped walnuts                                             nut
## 6 1 teaspoon salt                                                     salt
## 7 25 blueberries                                                      <NA>

The way case_when() works is just like a bunch of nested ifelse() statements. That is, if it satisfies a condition on the left of the ~ in the first line, it returns the output to the right, otherwise it goes to the next line. That results in “buttermilk” getting categorized as “butter”. If you wanted it to be captured as “milk” instead, you could switch the order of the butter and milk lines inside case_when().

I had to do this sort of thing a lot. For example, when “creamy peanut butter” was getting categorized as “cream” or “butter” instead of “nuts” or when “unsalted butter” was getting categorized as “salt”. You’ll also notice that if a description makes it all the way through the list, it gets categorized as NA. I’ll never be able to categorize all the cupcake/muffin ingredients, because people put weird shit in their baked goods.

# Conclusion

Webscraping can be frustrating, but you can set yourself up for success by choosing an easily scrapable website, annotating your code as you go, and taking measures to make your code as readable as possible. With big, messy data, you’ll likely never get it perfect, but you can use random samples of websites to help debug your code and test its effectiveness on new random samples of websites.

# Next Steps

Now that I have a data set I’m pretty happy with, the next step of the project is to do some exploratory data analysis to see what properties it has that are relevant to the sorts of multivariate data that ecologists deal with. Then on to statistical analyses to figure out what ingredients make cupcakes different from muffins. Is it sweetness? Is it something to do with leavening? Butter vs oil? Leave your predictions in the comments below!

###### Scientific Programmer & Educator

I’m a Scientific Programmer & Educator at the University of Arizona in the CCT Data Science group.