Schibsted aspires to be the leading media destination in the Nordics, reaching and empowering millions of people in their daily lives through offerings that are highly relevant.

As a media house comprising popular and diverse brands — including the primary news destination Aftonbladet, regional subscription newspaper Bergens Tidende, and niche news site Tek.no — users visiting any of these sites signal an interest in specific content, whether that is technology, politics, or e-sports.

However, until now, Schibsted has not had a uniform way of identifying users’ interests across the organisation. User interest predictions (UIP) was created to meet our ambition and ensure that we not only deliver valued personalised experiences to our users but also enable data-driven decision-making in the company.

Demographic data — such as age, gender, and location — have long been used as a proxy for interest. Other attempts at building interest profiles have focused on inferring interest by using newspaper sections.

Although there is overlap between section and interest, our findings showed that it is not readily interpretable as the same. Because Schibsted is a family of brands, each brand defines its sections independently based on its own needs, thus resulting in constraints created by brand-specific categorisations.

To solve the aforementioned problems, we leveraged AI. UIP is a machine learning-based solution that uses data from across all of Schibsted to predict users’ interests. Content is classified using natural language processing, then each user’s unique ratio of interest categories is aggregated based on their browsing history.

A fixed, well-known category taxonomy is used for content classification to ensure flexibility and sustainability for a fast-changing news landscape. This approach allows new interests to emerge as they become more relevant, and others can become dormant as their relevance decreases.

Expanding the model

In the first iteration of UIP, the model was developed to predict interest in 25 high-level categories — the top tier in our taxonomy.

UIP was later expanded to Tier 3 in our taxonomy, and the model can now infer interest in close to 1,000 categories. To democratise the insights this brings, interest dashboards have been created for employees to explore different combinations of brands, demographics, and more, uncovering the unique interest profiles of our reader base.

Aside from analytics, UIP has not only shown great promise for operational use but has also delivered outstanding results. For marketing purposes, the use of UIP, either alone or in combination with other models, significantly increases the click-to-open rate, engagement, and number of new sales compared to control groups and manual section-based interest categorisation.

Two UIP-empowered campaigns resulted in an astonishing 167% and 256% improvement in new sales. In advertising, UIP has enabled Schibsted to target new and previously undiscovered categories, and the sky is the limit to how UIP can be used for personalising content.

We sought to harness the power of being a family of brands and provide our users with personalised experiences based on their interests, not our guesswork. Our results prove that it is a winning recipe.

As more corners of the organisation awake to the power of interest-based personalisation, we are excited to see more novel uses of UIP.

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