How James (and Watson) learn to 'dance' with subscribers

May 25, 2019 at 05:21 am by Staff

A real estate agent's board as I make my way to News UK's office beside London Bridge offers, 'life isn't about waiting for the storm to pass, it's about learning to dance in the rain'.

Valid point: There's no doubt about the storm raging for newsmedia companies, and the damage being wrought by the smart algorithms of the likes of Facebook and Google. And no doubt about the need to learn.

Motivated in no small measure I'm sure, by threats of government intervention, Google is helping fund the learning that may help publishers dance. A million Euros (A1.62 million) of Google Digital News Initiative money has been channeled into a project shared by News Corp's Times and Sunday Times, and e-newsletter and edition specialist Twipe. And the one-time 'old lady of Printing House Square' is learning to dance. Automatically.

At the end of the year-long the project, I was fortunate to have the opportunity of checking on what had been achieved. Along with key staff from UK and European news publishers - among them The Guardian, the FT, Le Monde and NRC Handelsblad - we met on the seventeenth floor of News UK's headquarters overlooking the Shard and the rest of London, an inspirational setting which speaks of confidence and success.

What's interesting is not just the ability of AI learning to predict which stories a reader will find of interest, but what will engage him or her to the extent of maintaining a subscription they may have entered at a discount. Later it will be used to push registered readers - of which The Times has millions who joined to access a couple of stories - across the line to become subscribers.

This is the second and final major report card from the collaborative team which includes News UK data science and CRM specialists, and those of Belgium-based Twipe. You've read about the subject of their attention, quickly dubbed James, your digital butler.

He's been busy, as you'd expect, but the UK flagships of Rupert Murdoch's empire are in a hurry. The project came as News was already pushing visitors to its Times websites to share details such as their name and email address, latterly in exchange for free access to two articles, a move which almost doubled the number of its registered users. One of James' upcoming challenges will be to get these readers to leave their credit card, rather than just their name card, when they call.

A priority however has been to arrest churn by engaging subscribers in newsletter and edition-based products, and this is already delivering a return on the investment: Cancellations have been cut by 40 per cent among readers who interact with James - either by opening or clicking on the newsletter, which amounted to some 70 per cent of recipients.

Test and learn is a recurring theme, with AI used to check preferred time, format and frequency for the newsletters, predict content that will be popular and compare data alongside each subscriber's "propensity to cancel". With five recommendation models for timing, seven for content, 15 for format, and nine frequency propositions, the number of permutations is in the trillions, and the project team has turned to IBM's Watson to process the data at scale.

Each reader is however getting the product he or she wants - whether or not they know what that is - and the net outcome has been a 49 per cent reduction in number of cancellations by subscribers who interacted with James.

A core group of ten at Twipe in Leuven in Belgium and News UK's data science unit in London, have been supported by teams in Bangalore and Barcelona, so that a total of 25 are involved.

So far, the technology behind James (which stands for Journey Automated Messaging for higher Engagement through Self-learning) has been used to learn the preferences of 300,000 readers - including 117,000 subscribers - with each receiving content "on their own terms".

Learning what each reader likes, when (including how often) and in what format, at scale and "at cost" has been a massive task. Sometimes James has unearthed preferences readers don't realise they have themselves; sometimes more engaged readers have told News that they don't need help choosing which stories they want to read.

But it's the less-engaged group - making a sixth of the number of visits - which was the cause of concern, and their relationship with the mastheads has been strengthened and habits created.

Firstly, 70 per cent of the 117,000 subscribers in the trial interacted with James;

Secondly churn was reduced by 49 per cent when they did so.

Head of customer value Mike Migliore talks of the "nirvana" of engaged subscribers. "Machine learning reduces operational expectations all the time, putting the 'personal' back into personalization," he says.

During the project, 14 million emails were sent to subscribers and the various propositions "live evaluated". Just personalizing content was found to reduce churn by 14 per cent.

Project manager and Twipe data scientist Jasmien Lismont says individualizing the distribution involved learning what each reader wants, "including what they didn't know themselves".

There have been opt-outs - initially 15 per cent, including some who said they were quite capable of making choices for themselves - but this soon stabilized, and the more they interacted, the higher they engaged.

Other learnings included:

-the more articles "on the plate" the better - including 20 articles instead of five brought a 5.3 per cent better response, "and we haven't found the limit yet";

-a hybrid of trending and personalized articles worked 1.4 per cent better than just the most popular... and an editorially-chosen selection rated 2.4 per cent worse than this;

-a fixed sending time for each subscriber worked best, and the click rate was better for those opened soon after being sent.

The trial also explored 'next best action' as a way of extending reader interest in other ways, on-the-fly "exploration and exploitation" and context.

News senior computing planning manager Jackie Boholst also provided some insights into the support infrastructure, and IBM Watson allowing bespoke situations to be tested through the James algorithm. An API call initially looked to be the best option, but when numbers rose above 300,000 readers, a scalable database-led method won out.

She says the project will evolve and grow, with a new base template for emails, and James suited to being "the decision engine" for the future.

There's a neat irony in a project which will improve the health of and foster a greater direct relationship for a news publisher being largely funded by one of the tech giants most responsible for putting the industry under pressure, but we'll let that pass. Each has its own agenda.

What's encouraging is that this project has shown a fraction of what is possible for newsmedia companies, and hopefully the potential for the future. More immediately, Twipe expects to have beta products - including push notification - developed from the project, ready for release by the third quarter of this year... enabling many more to make the call, "Home James".

Peter Coleman

Sections: Digital business


or Register to post a comment