Gmail Priority Inbox & Explanation

Gmail Priority Inbox was launched a little while ago and is a nice system that attempts to learn to recognise the emails that are important to you and flag these as such - sort of reverse spam filtering.

The system uses a simple linear regression model to predict "importance" but customises the model, and the threshold used to get emails over the line into a user's priority inbox, based on a user's interaction with their mailbox. There's a nice paper about this by Google employees here.

The other nice thing I have noticed about this recently is that Google have included some nice explanation features - so when you look at an email you are told why it was classified as important. An example is shown below.

This is a nice example of an attempt to allow users look "inside" a machine learning-based application. It does, however, raise an interesting question. Given that users can now access information as to why particular classifications are made, what can they do about it? In the case of Google Priority Inbox, at the moment nothing. While we can tell the system that emails are important or not we cannot directly influence the decision making. This illustrates clearly the problem of balancing insight into the behaviour of machine learning-based applications with the resulting frustration users feel from not being able to change that behaviour.