Analysis of #FNCE tweets

The annual Academy of Nutrition and Dietetics’ (AND) Food & Nutrition Conference & Expo (FNCE) is a huge conference with an estimated 10,000+ in attendance, and there were many people people tweeting about it under the hashtag #FNCE. So I thought collecting the tweets & running some basic analyses might reveal some interesting things. If others want to poke around, I uploaded the dataset I collected here.

Methods

Unfortunately, I didn’t have the foresight to write a script to grab tweets directly from the #FNCE hashtag at the start of the conference, and because there is a limit to how many tweets you can backtrack on, I had to go about it from a very inefficient way. First, I wrote a python script to gather the twitter handles of those posting to the #FNCE hashtag starting at 0300 CST on Saturday, October 7th every 10 minutes through the end of the conference until October 10th (the conference took place October 6-9). This should represent the vast majority of tweeters, only missing those who may have tweeted or retweeted only before 0300 on Saturday and did not tweet for the rest of the conference. However, 40% of people posting to the #FNCE hashtag within the official conference start and end time did so only one time. I removed those that had 0 followers and accounts that I could not access (suspected spammers – there were 132(!)) using another python script. Finally, another script then read and searched each user’s last 400 tweets in their timeline for instances of the #FNCE hashtag. This took about 37 total hours to complete (had to scan over a half million tweets using 3 accounts to get additional API access, sorry twitter!). Tweet information and text containing the hashtag along with tweet ID, datetime, if it was a retweet, links, user bio, website, following and follower count, listed count, total tweet count, and location were added to a CSV file by this script, which I then performed some analysis with & made some graphs with in R.

Results

After restricting tweets to those posted after the official start of the conference up until the official end of the conference, there were 14512 total tweets (including old and new style retweets) from 1570 different people.

Here are the top 20 people who posted to the #FNCE hashtag (includes all retweets by the person):

Sore thumbs?

The mean, median, and standard deviation show that most actually did not tweet a lot:

Similarly, here are the top retweets, retweeted at least 10 times (an imperfect measure as retweets that add a message before the “RT” aren’t counted- this could probably be more precisely quantified by checking if a certain percentage of the tweet is the same as another). The top retweets are dominated by a small group of people:

There were more occurrences of “New-Style Retweets” (retweeting without adding the RT in front of the message) than “Old-Style”:

Because we have all tweets, if each tweet was an “old-style” or “new-style” retweet, and the dates for each, along with information like follower and friend count for each user, we could really do some complex analysis to visualize interactions. But that will have to wait for another time.

But since I collected the home locations of each person who tweeted with the hashtag (if they had one listed in their profile), I wrote a python script to geocode these into coordinates using the google maps API and plotted them on a map with R. Now we can visualize where people are from:

Here is only the US map:

Next I ran a check of frequent terms appearing at least 150 times. I did this for tweets that I stripped of usernames, hashtags, and links. These included all retweets. In alphabetical order:

[[1] “academy” “amazing” “awesome” “blog” “blogging” “booth” “breakfast” “calories” “check”
[10] “conference” “dairy” “day” “diet” “dietitians” “eat” “eating” “education” “energy”
[19] “excited” “exercise” “expo” “fat” “fnce” “food” “foods” “forward” “free”
[28] “fun” “getting” “grains” “health” “healthy” “help” “hope” “industry” “info”
[37] “kids” “krieger” “learn” “look” “looking” “loss” “love” “loved” “media”
[46] “meet” “meeting” “milk” “morning” “nutrition” “obesity” “people” “philly” “protein”
[55] “rdchat” “rds” “recipe” “reduce” “research” “session” “social” “stop” “sugar”
[64] “talk” “thank” “thanks” “time” “tips” “tweets” “twitter” “usa” “veggies”
[73] “weight” “world”

We can visualize associations between some of these terms. Lots of positivity sprinkled in there.

A word cloud isn’t that helpful in this case as most of the terms except for fnce are mentioned are similar frequencies:

We can dig deeper and examine correlations between specified terms. As I was following the tweet stream myself I noticed it was often very conflicted over the sponsors of the conference. I queried to see what terms were most associated with some relevant terms. The numbers under the terms represent the correlation with the queried term.

“sponsors”

“industry”

(Margo was a speaker at a relevant session)

“corporate”

The high association with “elated” is deceiving; in context it was a popular quote “elated” at critical tweets about corporate sponsors. The 3 digit numbers come from tweets/retweets of a session that included times and room numbers.

“healthy”

This only had 1 term that was correlated more than 20%:

And it appears to be from a popular tweet/retweet:

@Fooducate: “Healthy eating: education cannot by itself counteract convenience, pricing, culture, & marketing” #FNCE”

“GMO”

I was disappointed to read that there was an anti-GMO session (edit: “non GMO event” per tweets) (edit 2: this event was off-site and not part of the conference), so I was curious if this was reflected in the tweets. Indeed:

So it appears that there is a trend of dislike toward sponsorship, but this could be my added bias in selecting the terms. We can try a sentiment analysis of all tweets and see how positive or negative they are to add some objectivity. This scores each tweet by the number of positive or negative words it matches (positive or negative 1 per word) (match words from the opinion lexicon here, method based on this great example).

Finally we can plot the tweets by everyone over time to see the consistency of people joining in with their 1st tweet over time. As you can see, the people who tweeted earliest tended to tweet most frequently.

R is powerful but if you are new to it like me, it can be difficult to get it to work. Without the numerous examples shared by bloggers I would not have been able to complete these so quickly (or most likely at all). Special thanks to the following examples: wordclouds, associations, visualizing frequency, geocoding, sentiment.

Please let me know if the comments if you have any ideas or requests for additional things to look for (or tips!).