14 comments

[ 3.1 ms ] story [ 40.9 ms ] thread
Good to know, that before people enter a relationship, they interact, and afterwards, they express happiness.

I await further research predicting breakups, divorces, and stalker homicides.

+ elopements,adultery and the like.
Well, they can predict the likelihood of some of that: even in the same series the Atlantic is referring to, they gave some insights: https://www.facebook.com/notes/facebook-data-science/when-lo... (Adrien, the author, is a friend, but I haven't talked about that with him).

They can, but I’m assuming (like the rest) it's neither very new (unanswered messages, asymmetric attention) or very positive, or reassuring: I remember a blog post detailing how older, more male profiles spent time on younger, more female profile a lot, especially through photos. On Valentine’s day, big corporation have a duty to smile.

On every other day of the week, the ethics of informing, or acting upon any of that seems murky at best.

> For each timeline interaction, we counted the proportion of words expressing positive emotions (like "love", "nice", "happy", etc.) minus the proportion of words expressing negative ones (like "hate", "hurt", "bad", etc.).

I've always been suspicious of this type of measurement. It seems like statuses such as "I LOVE shoveling snow SO MUCH but I will be happy when winter is over!", where positive language is used to convey negative emotions through irony and contrast, would completely destroy such ratings. Maybe there are AIs that understand irony?

I would probably naively think that sarcasm/irony goes both way, and that it's most likely to cancel out or represent a negligible portion of emotion.
No AI understands irony perfectly, and even the most basic criticism are hard (“Thank you for the hair in my soup!“, ”Good riddance TelCo!”) However, it is irony because it's a minority use; words change their meaning if you use them too much in a different sense. This makes comparing the amount of both corpuses a crude, but relevant approach.

For client work, I personally prefer to look at the length, and extend of vocabulary (Zipf rank, excluding misspellings, argumentative keywords) of the comments, because it's not really positive vs. negative as much as useless (praises or attacks) vs. arguments (usable criticism vs. structured defense) that matters on the long run.

For relationship studies, I’d rather look into private messages and delay to respond (with, in Facebook case, a clear explanation that Scientist had no access to actual information, or even individual meta-data) than content.

What you have to keep in mind in cases like this is that sentiment analysis algorithms (such as the one you mentioned) or even machine learning algorithms in general are always playing the numbers. They usually* try to have the simplest model that can achieve a sufficiently high accuracy on a large dataset.
Reminds me somewhat of the old okcupid posts on statistics of love, attraction, dating, etc.
I really enjoyed their posts. I wish there were more of them / current versions.

http://blog.okcupid.com/

From what I understand, their blog was so successful that the team sold their expertise, and have evolved away from dating data. A shame, really: OkCupid could update their service -- but a great ressource for impertinent, if not highly un-PC tone, relevant if not brilliant methodology.
This shows as much about the data Facebook has as it does about how poorly Facebook works as a blogging platform.