#buzzfeed headlines

minimaxir:

Bonus Wordcloud of the Relative Frequency of each 3-Word Phrase

Tool is R/ggplot2. Data is more complicated and requires more explanation.

1) I used a scraper to get BuzzFeed article metadata (title, date, FB shares, etc.) for all ~69,000 articles and stored it all in a database table.

2) I decomposed each article title into its component n-grams and stored each n-gram as a seperate row in another database table (the table looks something like this). During the process, if a 1st or 2nd word in a title was a number (indicating a listicle), it was converted into a [X] in order to preserve and compare syntax.

3) JOINed the n-gram data with the article metadata, allowing me to aggregate phrases on any metadata field. (I limited the analysis to where number of occurences >= 50 in order to get a reasonable standard error)

I choose 3-grams since they provided the most insight in my testing. (Google Sheet of 3-grams)

Statistical notes:

1) Despite filtering on # >= 50, the confidence interval of all phrases is extremely wide, which shows a lot of uncertainty about the average and shows that using a linkbait phrase is not a sure bet for virality. (the exception is "character are you," which has an incredibly high lower bound regardless and shows that Buzzfeed's idea to switch to quizzes is smart)

2) I did not remove any stop words in the phrases because in this case, it's relevant. (e.g. big difference between [X] things only, [X] things that, [X] things you)