Predictive analysis of emails enables smarter conversations

By April 28, 2015
prediction email

With a view to making email interaction at companies more productive, researchers at Yahoo Lab have developed a machine-learning algorithm capable of predicting how an email conversation will run.

Having spent several months sifting through a database containing 16 billion emails exchanged by two million users, a team of researchers at Yahoo Lab have trained an algorithm to predict the behaviour of the users and so work out how to make a conversation more productive in the work situation. “With this algorithm, we can automatically establish a correlation between different factors,” claims Farshad Kooti, one of the research team, explaining: “For instance, if a young man receives an email in the morning, it’s very probable that his reply will be fast and concise. In fact we’re able to predict whether or not a given email message will receive a reply and if so how fast and how long the reply will be.”

A predictive algorithm

Kooti went on: “We observed the users’ behaviour – such aspects as the time it took to reply to an email and the length of the reply, plus the variables likely to affect this behaviour. Demographic factors such as age and gender matter but also the particular day and time when the email was received and whether or not it contained an attachment.”

Among the trends identified, the Yahoo Lab researchers’ report, entitled ‘Evolution of Conversations in the Age of E-mail Overload’, reveals that users tend to reply very fast to emails – in under two minutes in most cases – and that replies are in general very short, mostly less than twenty words. In addition, users tend to reply faster in the mornings and during office hours, but replies to messages that contain an attachment generally take longer than attachment-free emails. The youngest users tend to send shorter, faster replies than their elders. Similarly, men are faster and more terse than women. Lastly, when the number of emails received increases, the proportion receiving an answer decreases, which confirms the widely shared feeling that we are all drowning in a mass of emails nowadays and are unable to reply properly to all of them.

Towards smarter email boxes

The Yahoo Lab machine-learning algorithm could prove highly useful for email service providers: it could help them to work out a more efficient system than simply displaying by date and time order. "You can deduce how important a given email is for the user by the time s/he takes to reply to it. If the algorithm predicts that the user is going to want to reply to a given email very fast, this message ought to be placed at the top of the list of unread emails. Conversely, those emails that will be replied to more slowly – if at all – could be placed right at the bottom of the pile. This would reduce the chances of letting an important message slip through the net as the user wouldn’t have to sort his/her emails in order of importance one by one,” underlines Farshad Kooti.

This predictive tool could also help to improve the user experience. As many people use the same email address for both their personal and work emails, the algorithm could learn to distinguish emails in these two categories and so give the user the option of ignoring office emails when relaxing at home, and vice-versa.  One might also be able to adjust the content and time of delivery of one’s email so as to maximise the chances of actually obtaining a reply.


Legal mentions © L’Atelier BNP Paribas