What’s the link between football and machine learning?
Ever thought about the recommended stories you see along the side of a news article? Who makes those recommendations, and how do they choose the stories that appear?
Recommendations are a handy way for users to find further articles to read. They also keep users on our sites for longer, which is good for us.
And while there’s value in editorial teams picking related stories that might be a good follow-up read, we have another approach that is more personalised and frees up our journalists’ time so they can write more articles.
Working with machine learning company Seldon, we’re able to automatically recommend articles based on those that a user — and users similar to them — has read previously.
It works a bit like Amazon’s “customers who bought this item also bought…” algorithm, or the way Facebook chooses what appears in your news feed.
We use a mix of topic relevancy, published time and popularity to show a personalised list of recommended reads. The result is more stories you’re likely to be interested in, with a skew towards breaking and trending news.
It sounds obvious, but learning that a user is interested in business and recommending business news to them instead of entertainment, works pretty well.
Or more granularly, if someone reads about rugby more than football, then we can use this information to populate ‘their’ Sports section with more rugby news and less football. It’s a better user experience all round.
And we have some specific applications of this technology.
We can often predict with a high degree of accuracy where a user lives or works based on the stories they read. On our regional titles like the Manchester Evening News, we can then recommend hyperlocal news to a large proportion of readers.
This means a user could see more targeted news from Salford, or Oldham, or Trafford — wherever is meaningful to them.
It turns out that adding this into the mix really works. In A/B tests, more users clicked on hyperlocal news stories, with statistical significance at a high confidence interval. And after Manchester, we repeated this success in many other cities of the UK, each time checking the data validated our hypothesis before rolling out.
We also used behavioural data to improve our reputation amongst sports fans.
Some readers dislike seeing recommendations for rival football teams. So once a user has read enough stories about one particular team, we can say with some certainty they are are a supporter of that club.
We then recommend more stories about their team — and specifically exclude stories about certain local rivals.
We trialled this algorithm to improve our reputation and avoid negative user reactions to news about competing teams — but we found that the change also improved click through rates.
Up to 10% of users click on a recommended link to find their next story to read. So it’s a win-win.
We have lots more product development in the pipeline which takes behavioural data and uses it to improve the user experience in different ways across our platforms. Football is only the start…