Identifying depression on social media

I was skimming Arxiv yesterday and came across Moin Nadeem‘s paper on Identifying Depression on Twitter.

I am a little skeptical of the area due to the Samaritan’s Radar application. There is an awful lot that is good to the paper and the work in it and clearly the model has something very positive about it.

My concern, and I don’t see it addressed in this paper, is that it is a snapshot moment and may miss a trend.I think that if an application is going to look for depressive trends, or someone falling into a cycle, then it needs to be temporal as well. Instead of accepting that a Bag of Words at a moment is a key value in identifying an issue, we perhaps ought to see it in cycles and for applications to be able to model whether a social entity is getting worse or better.

That might make the application more useful but it does involve privacy issues. Therein lies a knotty problem. Although the data may be openly available, matching and following the data allows inference to determine other things within it.

That is not a circle that is easily squared.