Despite the proliferation of countless new modes of communication, email remains a centerpiece of many people’s online lives. Machine learning has the potential to remake how we interact with email, and there are several examples of this already in action today.
Help Drafting Emails
If you are one of the many million Gmail users, you are probably already familiar with one way machine learning is (hopefully) improving your email experience. Gmail’s “Smart Compose” is a relatively new feature that leverages machine learning technology to predict what you may type next. However, beyond the traditional auto-complete many of us have become accustomed to (which also relies on, in part, machine learning), Smart Compose will not just give you several options as your likely next word, but it will predict a single string of text that can sometimes be dozens of words long, allowing you to simply accept the prediction and move on.
This feature has the potential to save enormous amounts of time that would otherwise be spent typing out every word in every sentence-- and it is using machine learning to offer those predictions to you. By studying vast quantities of text, Google is using what you have typed so far to predict what you are likely to type next, given all that historic context about syntax, subject matter, and other language conventions. As users accept or reject the predictions offered, it can become more and more effective at learning the patterns that allow it to make more accurate predictions.
Help Prioritizing Emails
For many email super users, prioritization is the name of the game. Many people have created all sorts of rules and systems to help them quickly sift through tons of new emails, surface the most urgent and important, and take other steps to triage the inbound deluge. Machine learning presents the possibility of allowing your inbox to learn from your interactions and email to start to predict which emails are likely of greatest concern to you.
One strategy would be to build a regression model that predicted whether an email was likely important, using your past decision to engage with an email urgently. For instance, a model could be developed by looking at all a user’s past email history, and considering an email “important” whenever the user opened an email within a certain number of hours of receiving it, and then replying to it within a certain amount of time after that. Whatever the criteria, that would be used to allow the machine to learn about patterns in the sender, email content, and other information about the email to predict whether you might respond similarly for future incoming emails, and then flag those as important upon arrival.
A similar approach would be a classification strategy. Rather than predicting a single outcome (i.e., the email is important), machine learning could be used to group emails into distinct, predetermined categories. (Perhaps an unsupervised learning model could be used first to cluster all prior emails to help determine what these categories ought to be.)
Some tools, such as Superhuman, have already begun to deploy this sort of machine learning into their email offerings and have received positive reviews about it. Superhuman describes this as “anti-spam”-- essentially using the same techniques used to identify good emails, as we have learned to do to find and filter out bad ones.
Help Sending Emails
Another challenging aspect of emailing is the waiting game… waiting on a reply. One method to improve the rate of replies to an email is to make sure the email is sent at a time that is opportune for the recipient. But when is that? What makes for an opportune time? Luckily, with machine learning, we don’t need to answer that question generically. Instead, we can use tons of email data to pick up on subtle patterns that predict the likelihood of an email receiving a response.
Tools such as HubSpot are already including options for users to send non-time-sensitive emails to auto-send an email when it is most likely to be immediately opened and replied to. Machine learning makes that possible and potentially highly personal. It can look at the relationships between the content of the email, the subject line, and other patterns of sending, opening, and replying to help us understand how send time impacts the likelihood of a reply.
Of course, the timing of a send is likely just one variable that impacts the likelihood of a speedy reply, and so we can imagine a host of other machine learning powered features that could also make predictions (and, therefore, recommendations) about changes to email content, attachments, etc., that might improve our chances of hearing back from our recipient more quickly.
Why email is a good playground for machine learning
There is a lot about email that makes it an attractive place to put machine learning to work. First, there is a lot of data, for example: email address domains, times emails were sent and received, when and how often an email is opened, how much time is spent on an email, whether and how often an email is replied to, whether an email has an attachment, etc. Depending on what we are trying to teach a machine, this variety and volume of data can be a really helpful starting place.