Personalization of Local News Outlets

Sammanfattning

Ranking articles by predicted engagement is increasingly common in local news, but it creates an editorial risk; content that matters civically, such as coverage of public services, safety, and local government, rarely attracts strong click signals and often gets buried. In local journalism, where this kind of reporting is central to the outlet's role, the consequences are particularly serious. This thesis investigates how a local news feed can be personalized without losing civic coverage. To address this, a feed-ranking system was built and evaluated on interaction data from a regional news outlet. It combines two personalization approaches, one that matches articles to a reader's own history and one that draws on patterns from readers with similar interests. A rule-based layer on top enforces a configurable minimum number of civically important articles in each reader's top results. Evaluation used standard ranking metrics alongside measures of civic exposure, civic miss rate, and section diversity, making the trade-off between relevance and editorial goals observable rather than hidden behind a single accuracy score. The personalized system delivered 40\% better recommendations than a simple time-ordered feed, but left unconstrained, roughly one in eight readers received no civically important content in their top ten. Enforcing a minimum of one civic article per feed kept recommendation quality 24\% below the unconstrained system, though the constrained system still outperformed the time-ordered feed on both dimensions. The results suggest that personalization and civic coverage are not fundamentally in conflict: a well-constrained system can deliver relevance gains without abandoning the editorial obligations that motivate local journalism in the first place.