A few days ago, a PDF showing how Google’s Priority Inbox feature works circulated among the Hacker News and email marketing communities. The paper shows how the future of the web is evolving to deliver hyper-personalized results to users while relying on a huge sample of people connected through the cloud.
Priority Inbox, which attempts to deliver the most relevant emails to the top of a user’s inbox screen, combines the behaviors of all Gmail users with your personal preferences and behaviors to deliver an inbox where your most important mail gets read first.
It’s as if your doctor could compare your physical complaints with all of the symptoms experienced by people everywhere in the world, in order to deliver a diagnosis in a few seconds. Not impressive if you have a cold, but if you suffer from a rare disorder, it’s amazing. The paper is chock full of math and explanations of how Google (s goog) does this at scale (sharding databases and using Bigtable across tens of thousands of servers), but the crux of the matter is Google trying to apply machine learning to determine what each Priority Inbox user cares most about. To do that requires a computer connected to the cloud, and Google’s back-end servers. It’s an illustration of how massive computing power in the cloud and a client device can interact in ways that benefit users.
The biggest challenge isn’t necessarily the huge data crunching on the back end; it’s accounting for what the paper’s authors and statisticians call “noise,” and what I call the oh-so-human tendency to do what we want, not what’s most productive. For example, in email, we waste a lot of time and productivity opening silly emails about Lindsay Lohan’s latest escapades while ignoring those from our boss:
Opening a mail is a strong signal of importance for our metric, but many users open a lot of mail that is “interesting” rather than “important”. Also, unlike spam classification, users do not agree on the cost of a false positive versus a false negative. Our experience showed a huge variation between user preferences for volume of important mail, which can not be correlated with their actions.”
The challenge for machine learning is to calculate the signal from the noise on a massive scale in real-time, so your LiLo emails get sent to the bottom of the stack, but can still be read.
The researchers say that for Googlers who receive similar volumes of mail, Priority Inbox users spend 6 percent less time reading mail overall, and 13 percent less time reading unimportant mail. So while Priority Inbox may end up making you more productive, you might have less to chat about at the virtual water cooler. Unless you use those time savings to hang out on Twitter.
Do you use Priority Inbox? Do you think it makes you more productive?
Related content from GigaOM Pro (sub. req.):