Netflix Fails Data Anonymization

According to this story from the wired threat level blog, Netflix has shut down the sequel to its original $1,000,000 Netflix prize as a result of a privacy lawsuit. The problem for Netflix was that there is a specific law which prevents disclosure of a person's video rentals, and Netflix provided enough information about individual users in their supposedly anonymized training data that at least some of that data could be de-anonymized.

So, was Netflix wrong to give out the data it included in the second contest? Well, the second contest indicated what movies people had watched, and what ratings they had been given. The people weren't identified by name, but their ZIP codes, ages and gender, were provided. As it happens, there is an 87% chance that, if you have someone's birth date, zip code, and gender, you can uniquely identify that person (as related in this article, also from threat level). Does that mean Netflix's second contest ran afoul of the law? Well, it was settled, so we don't know what a court will say. However, it was certainly a significant enough risk that Netflix decided to cancel the well-publicized sequel to its earlier successful efforts, which probably means that Netflix made a bit too much public.

Now that it's all over, given the benefit of 20/20 hindsight, what should Netflix have done with the second contest? Well, from a conservative standpoint, it could probably have avoided the type of privacy complaints that came up if, instead of just removing names, it had followed the anonymization guidelines provided for medical research on human subjects (a good summary of which can be found here). That has the benefit of being the gold standard for data anonymization, and also including specific items to exclude, including the zip codes included in Netflix's data set.