Can we (yet) rely on computer algorithms and machine-learning to make sense of that most unpredictable substance in the universe – human behaviour?
“With a sufficiently large population,
the improbable becomes probable”
This week Microsoft has released a new tool to help you declutter your Outllook inbox. Called Clutter (side note: only available for Office 365 subscribers), it will analyse your inbox behaviour and then predict which emails you are more likely to read and which emails you are most likely to ignore. The latter will be de-emphasised from view to help you focus on what matters.
…or at least, that’s the theory.
We are experiencing a fresh round of virtual assistants, using technology to help us become more productive. There is nothing new with the concept, but modern tools are demonstrating a significant advance in how technology can help with everyday tasks.
The current trends can be divided into three categories:
Voice answering tools like Apple’s Siri, Google Voice and Microsoft’s Cortana are reactive virtual assistants. Ask them a question and they’ll respond with their best answer. The quality of those answers can vary significantly but significant progress has been made in recent years in being able to interpret natural language. Instead of having to do your best impression of HAL to be understood, you can vocalise more human thoughts: ‘Do I need a coat today?’
With notification centres like Google Cards and the Apple Notification Center, the clue is in the title. Instead of waiting to be asked, they will notify you of what’s happening. Some of the notifications are simplistic and well-established – new emails and upcoming appointments. But they are becoming increasingly advanced. For example, alerting you of about a traffic delay that might mean you ought to hurry and leave your home earlier than planned to start a journey if you don’t want to be late for a meeting.
The most advanced development takes those notifications a step further. Rather than prompting you to act on the notification, new tools will perform that action for you. Possibly the most well known is the Nest device, that will monitor your behaviour in your home in order to predict and optimise your heating. It includes a motion sensor and can automatically reduce or increase the temperature if there is motion (or lack of it) outside the normal patterns.
This third category is where Clutter fits. It makes a decision and acts on your behalf, deciding what emails probably matter and what emails probably don’t.
But the risk in relying too much on an algorithm to organise your inbox is in the potential severity of errors. Whilst humans rarely behave randomly, our patterns are not that easy to predict either. If they were, those annoying targeted advert systems that try to determine what ads to display around content on web pages based on your viewing history would be far less annoying and slow to spot behavioural changes.
Imagine the following scenario:
You have a distant relative who is in the habit of sharing pointless emails, be they chain letters, toilet humour or a particularly annoying obsession with cats on the Internet. They regularly share their habit with you and you pretty much ignore it all. The titles alone are usually enough to send the unread email straight to File 13 (aka the trash can). Clutter would correctly pick up this noise and dispose of it for you.
But one day they send something important, like announcing their impending wedding. Would Clutter interpret the content and spot the difference in tone? Who knows. Let’s try something a bit more vague. In reality you don’t entirely ignore your distant relative’s feline fascination. Your disciplined brain has several levels of triage: 1) Is the title pants? Yes = go to File 13/ignore, No = go to 2. 2) Skim the preview/abstract – Rubbish? Yes = go to File 13/ignore, No = go to 3. 3) Actually start reading the email – Bored within 10 seconds?? Yes = go to File 13, No = go to 4. 4) Failed to reach the end? Yes = go to File 13, No = go to 5. 5) Is it worth a response? Yes = send a reply, No = go to File 13.
Behaviour can vary at each stage and be influenced by other factors like time and mood. Depending on your preferred method of email management, the first two stages may often lead to the email being left in your inbox unread and ignored – you move on but don’t bother to delete (not everybody is addicted to the ‘inbox zero’ dogma). But other times you may decide you do delete them. And some days, you may decide you’ve got some time and feel in the mood to chuckle at the antics of cute kittens. Why the change in habit? You’re human. It would be boring if we actually behaved like the computers we try to get to do our work for us.
How does Clutter figure all that out? It will be interesting to see the accuracy observed at scale. Because the larger the data being analysed, the more likely that false positives will occur and the wrong emails will be classified as clutter
Machine-learning is a hot topic but we are still far from being confident about its accuracy when used to predict human behaviour. Recent history has demonstrated the pitfalls of relying too much on algorithms
For an excellent and pragmatic read about the potential and short-comings of ‘big data’, I highly recommend a recent interview with machine-learning expert Michael Jordan published by IEEE Spectrum, see the references below. Got to love the sub-title:
Big-data boondoggles and brain-inspired chips are just two of the things we’re really getting wrong
References
- De-clutter your inbox in Office 365 – Microsoft blog
- Machine-Learning Maestro Michael Jordan on the Delusions of Big Data and Other Huge Engineering Efforts – IEEE Spectrum
Featured image: ‘I love clutter‘ kindly shared on Flickr by Karl Sinfield
Joining Dots 