VRT MYNWS 2.0: insights from a pilot study on news personalisation


In June, the final pilot on news personalisation within the European innovation project CPN (Content Personalisation Network) was successfully finalised. With the project, we respond to the increasing interest in recommenders to select and personalise news articles. These recommenders, or algorithms, can be found in different web applications, from YouTube to Amazon, and often help achieve commercial goals. With the pilot study of VRT MYNWS, a new app, we aimed to personalise news in a transparant way and examine how recommenders could keep users better informed.

The pilot study

At the end of February, VRT Innovation launched a call for volunteers through the VRT Pilot Zone to test VRT MYNWS, an app that offers news articles based on your interests. 1500 people registered to test the app, of which a hundred people actually used it during several weeks to months. An overview of the results can be found below the article.

There are various algorithms that allow the VRT MYNWS app to personalise news articles. The algorithms used are: 

  • Content-based: First, the app registers which articles, images or audio users open and uses artificial intelligence to select similar content. Did you frequently visit regional news or political pieces? Then you can expect more suggestions that fall within these categories. Even if your reading behaviour changes, the software of VRT MYNWS automatically updates your news offering.

  • Collaborative: This algorithm allows the software to group users based on what they are reading, watching and listening to. As such, users are assigned a profile that shows specific similarities with others and will offer news articles that readers with similar interests will also find interesting.

  • Hybride: By using a hybrid approach, the recommender will combine results from other recommenders to get the best outcome from different algorithms. 

  • Advanced monitoring: This algorithm personalises content by monitoring reading time and app usage. If you would read an article for two minutes and scroll all the way down, this article will probably be more relevant to you than when you close an article immediately after reading. Next, your usage of the app can also help determine which articles you are being offered. For example, if you would not use the app frequently, you would also see older, but not less relevant articles.

  • Combatting the filter bubble: Here, we respond to a much discussed consequence of algorithms: the so-called 'filter bubble'. This entails that recommenders could narrow the minds of users, as they could get isolated in an information flow of similar stories. During our study, we have worked hard to adjust algorithms and prevent this from happening.

During multiple phases, testers of the VRT MYNWS app were divided into groups, each testing the app with different algorithms. By monitoring the behaviour of the testers, for example by reviewing how long they were reading and scrolling, and sending surveys, we were able to determine which techniques were effective and which techniques were not.

The results

Some interesting takeways from the data are:

  1. A content-based recommender can lead to longer reading times and deeper scrolling through the article. 

  2. A content-based recommender can increase attention for long-tail articles. This means that articles that are generally not being read often can generate new interest if they are being offered to the right users.

  3. A hybrid recommender can increase the offer in articles about different subjects, which can help pierce the filter bubble.

  4. Advanced monitoring based on reading time and app usage did not help stimulate readers to read more articles about different subjects. 

  5. Content-based algorithms can help stimulate users to read more articles about different subjects. This does need to happen with moderation, as to not alienate the reader. 

The most important takeaways from surveys with testers from different groups:

  1. They did not feel better or worse informed. 

  2. They did not fear to miss out on news by the personalisation of it. 

  3. They did not have the feeling of being in a filter bubble. 

We look forward to working on the following: 

  1. We have learned that some users would like to have more control about their news offer. Here, we could provide users with the option to submit their own preferences and interests.

  2. It is important that personalised news is an extra feature in a news app, an not the only one. As such, users would still be able to follow the general categories, such as the headlines, which are curated by the news editors themselves.

  3. In this study, it appeared that the filter bubble offered less of a problem than initially thought, since testers with personalised news were similarly unaware of the filter bubble as testers without personalised news.

  4. We discovered that testers liked to be surprised and that offering diverse articles also pleased them more. 

Related projects