Increasingly, the algorithms built into social media and your cell phone apps are quietly probing your psyche. But can they really do it? Georgia Tech’s Munmun De Choudhury studies this very question and says yes, they can.
Song: Jenny Scheinman, A Ride With Polly Jean
BROOKE: Increasingly, the algorithms built into social media and your cell phone apps are quietly probing your psyche. But can they really do it? Georgia Tech’s Munmun De Choudhury studies this very question and says yes, they can. Using technology carbon-based shrinks don’t have.
De Choudhury: They can look at data from hundreds of thousands even millions of individuals and learn to identify what is an expected pattern and what is an anomalous pattern. This ability to look at patterns which are different can be very powerful in identifying some kind of problem in their lives.
BROOKE: One area you've explored is postpartum depression. Somewhere between 12-20 percent of mothers experience it. But about half the cases go undetected. Tell me about the experiment you did. You followed 165 mothers Facebook feeds.
De Choudhury: Some of whom probably were diagnosed with post-partum depression and others or not. We requested if they would to share their social media data with us. And we looked at the data on Facebook and constructed a variety of different attributes of their behavior. Whether they were expressing positive or negative emotions. what was the intensity of emotion that they were sharing.
BROOKE: Did you have certain key words programmed in. Are the just obvious like 'happy,' 'sad' ?
De Choudhury: They are simple as them, phrases, key words, exclamations, emoticons and those kind of attributes which was serve as sort of gold standards so are indicative of positive emotions, what are indicative of negative emotions.
BROOKE: I know that you discussed the use of what's kind of 'very low-level language queues' that you wouldn't imagine would tell you anything about a person, but they do. For instance, one of them was the use of first person pronouns
De Choudhury: There is literature in psychology which says that inward thinking most of the time, thinking about themselves, is a characteristic which is highly observed in individuals who are suffering from mental illness such as depression. When we right? and post ? social media we don't quite think about these kind of aspects. It's kind of comes naturally to our writing and our algorithm was picking up those kinds of low-level queues.
BROOKE: Another low-level queue was an abnormally low use of articles. We're talking about such words as 'in,' 'on,' 'above,' 'with...'
De Choudhury: People who suffer from depression pay less attention to their surrounding and their environment so what we found is that individuals who had high-levels of depression they were using of these kinds of words.
BROOKE: Instead of the saying 'the refrigerator is broken' they'd say 'refrigerator broken?'
De Choudhury: Or maybe they would just no talk about the refrigerator because they are mostly just talking about themselves. Another finding that we noticed is when does the person post. One of the aspects about individuals with depress is that they tend to be unusually active in the nighttime. We basically look at when is that they are hanging out on Twitter or Facebook. And we find that to be a powerful predictor as well.
BROOKE: And so in your study of post-partum mothers, what did you find?
De Choudhury: So we found that we could actually utilize pre-partum queues to predict new mother's risk to post-partum depression before even the date of their childbirth. And that was really fascinating because early diagnosis is possible and that can really benefit the treatment of these kinds of conditions.
BROOKE: Another one of the projects to determine emotion is in the suicide watch forum on Reddit. Their it isn't just the subject you're studying, but the interactions and that's because of the potentially toxic environment on some of these sites.
De Choudhury: We wanted to see - how did the community try to help individuals who are struggling pro-suicide or self-harm type of thoughts. Sometimes they could be individuals who pretend to be trying to help. However their essentially trolls in the community. So what that means is they derive pleasure out of reading about other people suffering from these kind of thoughts. They want to encourage individuals to commit suicide.
BROOKE: So here is where your pattern recognition algorithm comes in because these trolls aren't necessarily overtly destructive. You might recognize them as trolls unless you follow their postings from moment-to-moment even site to site and your algorithm.
De Choudhury: Yes. We are sort of still in the early phases of this project but that's the kind of algorithm we are building. You know the pattern recognition part of the algorithm can really help us try to see that. Maybe there are queues in the way these trolls post, or where they post.
BROOKE: Is there a question that you've put to your algorithms that they just can't answer.
De Choudhury: we are missing people who are actually depressed but our algorithms - based on the kind of emotion they use, the kind of social support they have - we're just not able to detect that they're depressed. So that's bad. But I think even worse, would be when they are actually not depressed but they're just not active on social networks. which is why our algorithm thinks that they don't have the right kind of social support. Or maybe they're just a generally a negative person but that might not mean that they're actually depressed.
BROOKE: Right, how do you deal with false positives and false negatives. It' interesting I tried this crude algorithm that looks at your Twitter postings and develops a quick character profile it's called "analyze words.' And they break it down - positive, negative, arrogant, personable, blah blah blah. And I came out WAY more positive than I actually am in real life. I think this is because I have a more upbeat Twitter personae. And how do you reckon with the fact people can present themselves in all sorts of ways.
De Choudhury: That's where sort of looking at many, many individuals helps in certain ways. Because that helps you set a baseline behavior - what to expect. That certainly helps in curbing some of these limitations.
BROOKE: Munmun, that you very much.
De Choudhury: Thank you.
BROOKE: Munmun De Choudhury is an Assistant Professor at the School of Interactive Computing at Georgia Tech. And a Faculty Associate with the Berkman Center for Internet and Society at Harvard.