A Data-Driven Historical Analysis of Cosmo

Using NLP and image processing to analyze trends on Cosmopolitan Magazine covers

Magazine covers have an underrated influence on our culture. Even if you don’t read tabloids or women’s magazines, you can picture a Cosmo or People magazine cover and can give examples of the types of stories they run. Every time you’ve purchased groceries or gotten a snack for the plane, you see these magazines and they make an impression. And as you wait in line and read the sensationalized headlines maybe, every once in a while, in a weak moment, you even purchase one (just me? Okay).

To try to further understand magazines’–specifically women’s magazines’–market appeal and impact as cultural influencers, I decided to use data analytical tools to abstract their content strategy and marketing techniques. I extracted the text from magazine covers, performed NLP topic modeling, and used image processing techniques to understand graphic trends and representation. I wanted to analyze how their intention to appeal to women migrated from this:

to this:

I entered into this investigation with a few big questions about methodology. First, does this data source exist? I wasn’t sure if anyone had digitally assembled all the covers of any women’s magazine. Luckily, the academics had me covered. There is a Proquest database called the Women’s Magazine Archive that has covers of various magazines, including Cosmo from 1925-2005. That explains how I landed on this specific publication. I ended up using covers from 1950-2005 from this database and individually searching for and downloading covers for 2006-2019.

My second question was “can I use computer vision techniques to record text off of the .jpg files?” This had a more complicated answer:

I used Google’s tesseract optical character recognition (OCR) engine, an open source tool trained to recognize and extract text. Tesseract worked well on covers up until the 90s, with a single background color and text color, because the process uses binarization, bringing the pixel values above a threshold to white, and below to black.

Later covers with more sophisticated graphics meant tesseract ocr’s default process lost more than half the text. To solve this, I inverted the images and changed the binarization threshold to be between the background color and the lighter text colors and so was able to extract a larger percentage of the text.

However, I used a standard pixel value for all covers as the binarization threshold, I think that this would have a higher retention rate with a dynamic threshold that varied based on the background and text colors, a space for further work.

My last question was “what information can I extract using data science processing techniques that go beyond what I can generalize just looking at some example covers?” Leveraging natural language processing, clustering algorithms, and composite images, the answer was, “Quite a lot!”

With the recorded cover text, I performed natural language processing. Using NMF topic modeling, I saw these 6 topics over the 70 years of covers:

Surprisingly, two out of six topics were associated with literature: fiction and mysteries, which seemed like a large percentage, based on my perception of Cosmo in my lifetime. However, I learned that the magazine serialized novels and short stories up into the 50s, and fiction continued to appear into the 90s.

As you can see on the graph below, which is of the percent of words on the cover in each topic averaged per year, the literature topics continuously decreased during the time period of covers I analyzed.

The vertical lines here denote changes in editors. Clearly, topics covered change when editors change. With the first female editor in 1965, Helen Gurley Brown, there is a decrease in literary topics, and in turn you get more about women’s individual lives. When Brown’s 20-year tenure ended. the shift away from literary content stepped up again to an extremely high percentage of words in the “sex” topic.

To understand more about these topics after the era of literature publication, I performed NMF on the text from 1980-2018 and found these topics.

If we look at a percentage of words in each of these topics on covers we can see that the 2000s showed a decrease in the “Relationships” topic reflecting a societal shift away from defining women via boyfriends and marriage.

At the same time, the health/beauty and sex categories rose, and we also see an increase in the percentage of “power” words, indicating an editorial shift toward empowerment, away from condescending topics like “how-to guides.”

Finally, I also did image analysis to further support my textual analysis. Using average pixel images I created a composite cover for each decade, and the 50s and 60s did not have a cohesive style (see the 1962 and 1964 images) besides the ever-present Cosmo logo. However, when I divided the composite into before and after 1965, I found that I could see a portrait start to appear, shifting to a single “cover girl.” This further supports my findings that Cosmo’s content in the late 60s shifted to focus on the woman as an individual and her particular interests.

Then starting in the 70s we see the standard cover consisting of a central woman usually white posing with 3/4 of her body shown.

With the more contemporary topic changes to more empowering language, I am interested to see how much longer this style persists, because to me it feels a bit more objectifying than empowering, and out of curiosity I looked at the most recent covers at the time of this analysis:

These are March and April’s covers from this year (2019). In one, we have a pizza front and center, and the other is the first cover with multiple people since the 1960s. It does seem the mold is starting to break down.

In using data to understand text and graphic changes, I was able to answer my final question:

Are magazines like Cosmo merely a reflection of changes in attitudes towards women over time, or are they driving the changes in those attitudes?

The answer is likely both. We can see the changes in Cosmo’s topics correlate highly with the different waves of feminism, and, likely, the magazine played a part in making sexual empowerment mainstream. But if society were not already moving in that direction, then the magazines wouldn’t sell. I think that they want to be seen as pushing the boundaries and empowering women, but not to a level that is close to the most progressive feminist thinkers.

Overall, my research convinced me that we are continuing to see growth and changes in society’s acceptance and support for the autonomy and individuality of women, and that these changes are reflected in the commercial reading material that is meant to attract and portray us.

Work It

I Put My Code Down, Flip It and Reverse It

I had limited experience with programming before starting the Metis Data Science boot camp a week ago. I majored in chemical engineering where I used MatLab for some classes analysis and took a beginning Java course. I definitely had never tried to write "good" code. If it ran, I was happy. I definitely never accounted for edge cases and I definitely never tried to make my code run quickly or make it legible.”

Quickly, the Metis instructors prioritized these qualities that make for "good" code. They, along with Missy Elliot, showed me that it is important, even when a questions is answered, to look at it from another angle to make it better, and sometimes even "reverse it."

We first had to write one of the classic beginner programs, "FizzBuzz." For those that have not been indoctrinated, this program prints out consecutive integers 1-100. For multiples of three, you print "Fizz", and multiples of five, you print "Buzz." If a number is a multiple of both three and five, you have to print "FizzBuzz." My partner and I soon ran into the problem that most beginners find: you can say "if divisible by three, print Fizz; if divisible by five, print Buzz," but then you’ve already taken care a number like 15 in the first statement. In this structure, 15 can never reach an opportunity to output "FizzBuzz."

We then took the two "if statements" and tried to extend them, thinking about string concatination. If it was divisible by three, print "Fizz." Take that same integer and, if it is divisible by five, add on "Buzz." Anyway, we messed with this for a while and we could get a "FizzBuzz" to print, but whatever constraints we tried to give, something would go wrong and it would override those that were just supposed to print "Fizz," or some other issue.

As you may know, the answer involves starting by checking if the integer is divisible by 15, then moving on to 3 and 5. No combining strings necessary:

1
2
3
4
5
6
7
8
9
for i in range(1,100):
  if i % 15 == 0:
    print(\"FizzBuzz\")
  elif i % 3 == 0:
    print(\"Fizz\")
  elif i % 5 == 0:
    print(\"Buzz\") 
  else:  
    print(i)

Immediately it was clear that we needed to start thinking "out of order." For whatever reason, starting with 15 before 3 was not where my brain wanted to go. This happened time and time again over the course of this week. We would want to loop a string or a list and it would make much more sense to start from the end. Or we would want to guess a number and it would make sense to start from the middle and work our way up or down. In any case, the very beginning was often not the best place to start. ”

This extends beyond the iterating through a list or creating the most efficient short snippets of code. When we were working on our first group project, I found myself consistantly running into situations where I would think a bit of code would run, and I would soon have a rude awakening. My conceptual understanding was often fine, but I had encoded something in an unexpected data type earlier on, or something like that. Then, since the beginning of my code was working well, I would instead try to work with this different data type to do a task that would be simple if it were in a different form. Why didn’t I just recode the beginning? Who knows, but eventually I extended this "reverse it" model to mean: sometimes the most "natural" order in which to work is definitely not the most efficient.