Collaborative Filtering for News Recommendations

When thinking about recommending news articles, people often start with the content of the articles themselves.

If someone is reading an article about climate policy, recommend another article about climate policy. If someone reads about football, recommend more football. If two articles mention the same keywords, names, or places, they must be related.

That sounds intuitive.
But in news, it is often not enough.

The reason is simple: readers do not only follow topics. They follow intent.

Two articles can contain many of the same words and still serve completely different audiences. One may be a serious business article about a media merger. Another may be a gossip article about the same people at an afterparty. A content model may see overlapping names. A reader sees a completely different reason to click.

That is why collaborative filtering is so powerful for news recommendation.

What Content-Based Recommendation Does

Content-based recommendation asks a straightforward question:

Which articles look similar to the articles this reader has read?

To answer that question, the system looks at signals such as topics, tags, entities, authors, categories, and words in the article. If an articles share many of those signals with the user's recent reads, the system considers them as most relevant.

This can be useful. If a reader is following a developing story, content similarity can help them find more articles about that story. It can also work well for clear topic preferences, such as football, finance, culture, or local politics.

But content similarity has a limitation. It does not know why the reader clicked.

The article may mention a politician, a company, a city, a celebrity, a court case, or a sports club. The text alone cannot always tell whether the reader is interested in policy, finance, entertainment, local relevance, human interest, or background context.

What Collaborative Filtering Does

Collaborative filtering asks a different question:

What did similar readers actually read next?

Instead of starting from article text, collaborative filtering starts from behaviour. It observes patterns across many readers: which articles they read, which paths they follow, which articles they ignore, and which pieces of journalism tend to belong together in real reading sessions.

The intuition is very human. If many readers who behave like me also read a certain article, there is a good chance I may value that article too.

This does not require the recommended article to look textually similar to the article I just read. It only requires a strong behavioural signal.

Similar Readers, Not Just Similar Articles

The most important shift is from article similarity to audience similarity. A content-based system says: this article looks like that article. A collaborative filtering system says: readers who liked these articles also tended to value that article.

That difference matters because readers rarely spend an entire session reading only one topic. Within a single visit, they may move from politics to local news, from business to consumer advice, or from a serious investigation to a lighter human-interest story. These reading journeys reveal connections that are invisible to content analysis. Behaviour captures not only what readers read, but also how they naturally move between different types of journalism.

Collaborative filtering therefore helps discover relationships that editors and content models may not explicitly define. It learns from the collective behaviour of the audience.

Same Words. Different Reader Intent.

Consider two articles that mention the same people, organisations, and locations. A content model may conclude that they are highly related because the names overlap. For a publisher, that can produce recommendations that feel technically correct but editorially wrong.

The reader who is interested in a serious business article may not want a gossip article simply because it mentions the same executives. The entertainment reader may not want the market analysis either. The words overlap, but the intent does not.

Behaviour makes that distinction visible.

Why This Matters for Publishers

Newsrooms produce many different kinds of journalism every day: breaking news, local reporting, investigations, analysis, interviews, service journalism, culture, entertainment, sports, and background explainers.

The goal of recommendation is not simply to show more of the same topic. It is to help each reader discover more of the journalism they are likely to value.

That requires more than matching words. It requires learning from how readers actually move through the publication.

This is also why collaborative filtering can increase discovery. It can surface articles that are not obviously similar in text, but are connected by real audience behaviour. It can help readers move naturally from one interest to another, instead of trapping them inside a narrow content category.

Content Still Matters

This does not mean content signals are useless.

In practice, strong recommendation systems combine multiple signals. Content can help with cold starts, breaking news, editorial rules, freshness, safety, topic constraints, and explainability. Behaviour can help identify what readers actually value once enough interactions are available.

The strongest systems do not ask whether content or behaviour is the only answer. They ask how to combine both in a way that supports the publisher's editorial strategy.

But for news recommendation, behaviour is often the signal that makes the difference. It captures reader intent in a way that article text alone cannot.

The Core Idea

Content-based recommendation starts from the article. Collaborative filtering starts from the reader. That is the essential difference.

For publishers, this distinction matters because journalism is not consumed as isolated pieces of text. It is consumed through habits, interests, moments, sessions, and reader journeys. Collaborative filtering helps reveal those journeys.

And when recommendations are based on what similar readers actually value, more journalism finds the audience it was written for.