Machine
learning: what our tech team wants you to know
Published
on June 21, 2018
“Machine learning
extracts the magic dust out of data.” – Ralf Herbrich, Director
of Machine Learning Science at Amazon
At the GEN Summit in Lisbon,
Benedict Evans called for us to change the way we talk about artificial intelligence. So when
members of our tech team at Twipe recently attended the AWS Summit in Berlin,
I thought it was the perfect time to see what they had to say about AI and
machine learning.
To begin with I asked for a
simple definition of machine learning, to make sure everyone was on the same
page.
“Machine
learning uses algorithm that gradually improve on a task without explicitly
being told how, i.e. they ‘learn’ from data.”
Joris Gielen, AI
& Software Engineer at Twipe
We also put together a
short guide on the different types of machine learning as well.
What can non-technical people do
with machine learning?
There are different levels of
what you can do with machine learning, and it’s important that all types of
profiles can access the systems. In fact, a good machine learning system
requires strong skills in three different areas: development, business, and
stats. It’s hard to find one person with all three skills, so different tools
have been developed to help users benefit from machine learning, no matter
their background. One such tool is SageMaker, a
fully managed platform that allows anyone to easily build, train, and deploy
machine learning models at any scale.
What do we need to consider when
working with machine learning?
There can be unintended
consequences with machine learning, as machines can learn things that we didn’t
foresee and they lack the intuition that humans have. An example of this is the
case of a robot which was told to learn to walk with the least contact with the
ground, as this is how cheetahs are able to run very fast. However the robot
learned to turn over and crawl on his back, as this way his feet didn’t touch
the ground at all. While he fulfilled the task he was asked to do, this wasn’t
what the intended outcome had been!
It is also important to
understand the tradeoff between efficiency and quality when developing an
algorithm, as the one with the best result might not be the most optimal for
your solution.
How will machine learning impact
the media industry?
Last year McKinsey released a
report on the impact of machine learning on the media industry, including the
most likely use cases. We’ll begin to see even more hyper-personalised
advertising, price and product offerings, and recommender systems. We’ll also
see journalists being able to quickly spot new trends in consumption patterns,
such as identifying viral content before it goes viral. Reducing reader churn
will become more of a data-driven process (something we’re working on at Twipe
now!).
What are newspapers doing with
machine learning today?
The New York Times announced this year they have
launched “nytDEMO“, a cross-functional team that will build data and
technology solutions for brands, using the same tools and insights that power
the newspaper itself. Their first machine learning project was “Project Feels”,
which looked at how reader emotional response to articles influenced
engagement. Now they’ve launched perspective targeting, which allows
advertisers to target their media against content predicted to evoke certian
reader sentiments, such as self-confidence or adventurousness.
Schibsted is also busy using machine
learning in a variety of ways. In the newsroom itself they’ve implemented
automatic tag suggestions for tagging articles, while for readers machine
learning is used to personalise the front page and the content they see on the
website.
Neue Zürcher Zeitung has built a flexible paywall using machine learning, so that is is
personalised to the individual based on hundreds of criteria. In the three
years since they built this, they’ve increased their conversion rate by
fivefold, with 2.5% of people who view the payment message becoming
subscribers. The system looks at data including reading history, device and
time of day to alter the paywall message. Looking at reading history helps to
better communicate the value a subscription would offer.
Of course, News UK is also working on a machine learning
project right now with us at Twipe. “JAMES, your digital butler” will use machine learning to
gradually get to know the habits, interests, and preferences of readers.
He will expose them to relevant content in editions–current and past–in
readers’ preferred formats, channels, times, and frequencies. This will
increase reader satisfaction and engagement and ultimately accelerate
subscription growth, enabling JAMES to transform conversion and engagement
strategies by moving from segmented to highly individualised interactions with
readers. To learn more about this project, make sure to attend ConTech in
London this November, where Twipe CEO Danny Lein will
be discussing our learnings from this project.
What is the future of machine
learning?
Every day, every minute even,
there is more and more data available to us, with the amount of data we’re able
to access only growing in the future. We don’t know yet what we don’t know
in regards to machine learning–as Benedict Evans explained at GEN Summit, no
one was thinking of car hailing apps like Uber when cellphones were first
developed.
“We’re
just at the start of what’s possible with machine learning.”
That’s why it is important now to
think about the scalability of our machine learning systems, so they’re able to
grow with the influx of data in the future.
How can I implement more machine
learning in our working processes now?
I also asked the team for their
advice to publishers wanting to work more with machine learning, here’s what
they had to say.
“I
would advise publishers to first start with a descriptive analytics project, so
taking a better look at your readers and understanding how they’re reading for
example. Then you can move on to larger projects, such as creating a predictive
model to predict churn.” – Jasmien Lismont
“Machine
learning might sound scary and difficult at first, but you don’t need to be
afraid. Decide what you want to extract from your data and give it a try. It
might not be easy at first but I guarantee you: you will learn a lot, and
you’ll quickly discover what works and what doesn’t.” – Bram
Hendrickx
“No
matter your technical background, you can succeed with machine learning if you
pick the right tools. In the beginning, make sure to choose tools that will
allow you to control the level of complexity.” – Lies
Tambeur
“My
advice to publishers wanting to work with machine learning is simple: focus on
your data. The first step is to collect lots of data, then you can work on
analysing and cleaning your data.” – Joris Gielen
Thanks to Joris Gielen, Jasmien
Lismont, Bram Hendrickx, and Lies Tambeur for answering some of the common
questions publishers have about machine learning.
Nenhum comentário:
Postar um comentário