Sweden’s VK Media has a long-term goal of supplying interesting and relevant articles to all readers — regardless of where they live or their interests.

A part of realising this goal is to develop a personalisation platform that can help us highlight certain content for the users interested in it. The ultimate goal is that no user misses out on any article that is relevant to them.

As the first step in this direction, we built a personalisation feature called the Powerpuff with the goal of improving the experience for super-local users on VK Media’s Västerbottens-Kuriren and Folkbladet sites.

Västerbottens-Kuriren and Folkbladet cover a total of 15 municipalities in the county of Västerbotten in northern Sweden. The smallest municipality has about 2% of the population size compared to the largest one, which makes it challenging to cover all these municipalities fairly and make all the great content easy to find.

Content from the smaller municipalities is easily lost in the news feed, which means the super-local subscribers have a harder time finding it and might even miss out on it. We realised this might be the perfect case to try to improve with our new personalisation platform.

The foundation of our personalisation engine is a data platform, where we collect behavioural data from our sites and metadata about our news content. This means we can analyse what our logged-in users read and seem interested in.

For Powerpuff, we set up a machine learning model to find interest clusters on the municipalities we cover. The resulting clusters showed us which (if any) municipalities each user seemed interested in.

When working on the model, we realised the data needed to be weighted in various ways. For example, smaller municipalities needed a boost compared to larger ones, since there is much less content about them. No matter how interested a user is in a subject, they can’t read more about it (on our sites) than we publish.

The news values (i.e. how “big” the story is for the general population) set by the newsrooms also proved important — users reading articles with low news values indicate a high interest in that subject, while reading articles with a high general news value might not tell us as much about that user’s specific interests.

Data-based decisions

After having found these municipality clusters, we built the Powerpuff. The project was a joint effort between several different teams (development, data, UX, and the newsrooms). It was an iterative process in which we developed new features and evaluated them continuously.

We designed the Powerpuff as a special teaser on the first pages: it looks like the normal “puffs” (teasers), but contains a special recommendation for the logged-in users interested in a certain municipality.

To evaluate the performance of the Powerpuff, our main measure was the CTR (click-through rate), and we continuously made adjustments to avoid cases with lower CTR.

For example, we noticed many users had already read the recommended articles. When we started to only recommend newly published articles, we saw an instant increase in the CTR.

The main thing we realised, however, was the difference in CTR between the largest municipality and the smaller ones. It turned out the need for special highlighting of news from specific municipalities did not include the largest municipality, since most of the main feed already covers it.

When we adjusted so that only users interested in smaller municipalities would get these recommendations, the total CTR increased greatly. For these users, the Powerpuff performed better than a normal teaser at a comparable location on the site. The CTR for the Powerpuff was approximately 18% higher at vk.se and about 8% higher at folkbladet.nu.

This first step for our personalisation platform showed us that our data can help us identify our users’ interests. If we find the right ways to use this, we’ll be able to recommend even more relevant content. In this first case, the key was to combine insights from the data with our knowledge of the local area.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .