Accountability in Data Mining: A new seminar

Glencora Borradaile makes a number of interesting points about the implications of Aaron Swartz's life and work for us academic computer scientists. 

As computer science academics we are in a very powerful position. We are trusted with shaping the next generation that will make very important decisions that will have far-reaching social implications. Decisions like those over Facebook’s privacy defaults, motivating technology that enables autonomous private vehicles at the expense of the public interest, defining ownership of electronic media. We make those decisions ourselves in our research; what we research, how we allow our research to be used.

 Whether or not you agree with her particular policy preferences, the point remains that the technologies we develop can have lasting consequences for the "shape" of the world to come. And if we don't speak up (either through our work, or through our own advocacy), others will take the technologies we develop and find their own ways of using or misusing them.

All of this leads up to my current interests in data mining. I've been thinking about the problems of accountability in data mining for a while now: initially mostly in private, but now more and more in public (along with +Graham Cormode and +Andrew McGregor) as I see the importance of the issue.

What is accountability in data mining ? It means many things really, but for me, for now, it means this:

The fruits of data mining pervade every aspect of our life. We are recommended books and movies; given differential pricing for insurance; screened for potential terror threats; diagnosed with various diseases; and targeted for political advertising. The ability to sift through massive data sets with sophisticated algorithms has resulted in applications with impressive predictive power. 

Since the internals of a learning process can be complex and opaque, it is important to verify that the results satisfy the properties claimed by the algorithm. The importance of this goes beyond merely checking the algorithm itself. Validation mechanisms also provide a way for users to demand accountability from authorities who might use the results of mining to affect users in some way (by changing their insurance rates, or even putting them on a terrorism watchlist). As the results of data mining affect more and more of our lives, the more crucial it is that the user be able to validate decisions made on their behalf and that affect them. 

The above is an introduction to a seminar I'm running this semester on this topic. I'm a little nervous about it, because the topic is vast and unstructured, and almost anything I see nowadays on data mining appears to be "in scope". I encourage you to check out the outline, and comment on topics you think might be missing, or on other things worth covering. Given that it's a 1-credit seminar that meets once a week, I obviously can't cover everything I'd like, but I'd like to flesh out the readings with related work that people can peruse later.

It's entirely possible that we don't care about our privacy any more (as Facebook seems to think*). But we will care about the consequences of the use of our data. My perspective is that a better understanding of what is technically possible (and not possible) to demand and get accountability will be critical to informing larger policy discussions on this topic.

* In an earlier version of this post I wrongly attributed this sentiment to an individual. Apologies.