Summary of News Recommendations

Overview of News Recommendations

This episode of Data Skeptic focuses on news recommender systems and why they are fundamentally different from recommending movies, music, or products. Kyle Polich speaks with Andrea Ayana, a postdoctoral researcher at the University of Mannheim, about her work on neural news recommendation, responsible and inclusive AI, and her open-source framework NewsRecLib. The conversation explores the technical challenges of news personalization, the societal risks of filter bubbles, the limits of current benchmarks, and why simpler models can sometimes outperform more complex ones.

Why News Recommendations Are Different

News recommendation is not just “movie recommendation with articles”:

• News is time-sensitive: relevance decays quickly, unlike evergreen content like films.
• Feedback is weaker and implicit: users rarely rate articles; systems mostly rely on:
  • clicks
  • dwell time
  • session behavior
• The stakes are higher:
  • Poor movie recommendations waste time.
  • Poor news recommendations can reinforce bias, polarize users, or trap them in filter bubbles.
• News affects democracy: recommendation systems shape what people are exposed to and can influence worldview.

Responsible and Inclusive AI in News Access

Andrea frames responsible and inclusive AI as systems that optimize for more than just engagement or click-through rate.

Key values she highlights

• Diversity of viewpoints
• User inclusion
• Linguistic inclusivity
• Reducing bias amplification

Why this matters

• Recommenders can subtly shape opinions by filtering out information that does not match prior preferences.
• This can lead to:
  • filter bubbles
  • opinion polarization
  • radicalization over time
• Systems should support not only dominant languages like English, but also low-resource and underrepresented languages.

Language and Multilingual Challenges

A major theme of the interview is the gap between current AI capabilities and global language diversity.

Main points

• Large language models still support only a fraction of the world’s 7,000+ languages.
• Many languages are:
  • low-resource
  • linguistically distant from English
  • primarily spoken rather than written
• Performance drops sharply for languages far from English.
• Machine translation is improving, but it still struggles with:
  • domain-specific terms
  • idioms and expressions
  • spoken-only or non-standardized languages

Takeaway

Andrea argues that multilingual systems need special care, not just “more time” for current models to naturally solve the problem.

What Is NewsRecLib?

NewsRecLib is Andrea’s open-source framework designed to make news recommendation research more reproducible and comparable.

Purpose of the library

• Centralize many news recommendation models in one place
• Standardize evaluation
• Make comparisons more fair and reproducible
• Allow researchers to:
  • mix and match model components
  • plug in their own models
  • test on shared datasets

Why it was needed

Andrea found it difficult to compare models across papers because:

• every group implemented systems differently
• many benchmarks were not directly comparable
• reproducibility was poor

How Neural News Recommenders Work

The conversation breaks down a typical neural news recommender into two main parts:

1. News encoder

• Takes article content, often the title/headline
• Produces a numerical representation of the article
• Can also use metadata like:
  • category
  • entities
  • viewpoint/sentiment

2. User encoder

• Aggregates clicked articles into a user representation
• Uses clicks as a proxy for interest
• Then ranks candidate news articles by relevance to that profile

What Actually Drives Performance?

One of the most surprising findings in Andrea’s work is that bigger and more complex is not always better.

Findings

• The news encoder matters most for ranking performance.
• The user encoder often contributes less than expected.
• In many cases, simple user aggregation methods perform as well as more complex approaches.

Why?

Andrea suggests the reason is mostly data limitations:

• news interactions are short-lived
• click data is noisy
• dwell time is often unavailable
• users are frequently not logged in, so long-term history is limited

Practical implication

Researchers should not automatically assume that adding complexity improves results.

News Content Modeling: What Helps Most

The episode also discusses what improves article representations.

Earlier approaches

• Used static word embeddings, which had limitations:
  • no context
  • ambiguous words like “bank” share the same representation regardless of meaning

Additional signals that helped

• article category
• named entities
• metadata

Modern approaches

• Pretrained language models now capture much of this context directly.
• Larger LLM-based encoders can help, but smaller models can still be very effective.

Are Current Benchmarks Good Enough?

Andrea is skeptical that the community currently has the right evaluation benchmarks.

Problems with existing benchmarks

• Many are old
• News changes rapidly, so datasets become stale
• They focus heavily on text only
• They often ignore:
  • images
  • video
  • multimodal cues
  • modern consumption habits like short-form video news

Dataset concerns

• Open news datasets are limited because media companies rarely release data publicly
• Common datasets include:
  • MIND (Microsoft News Dataset)
  • EbNERD (a newer Danish dataset)
• Datasets vary a lot in content:
  • MIND contains more “soft news” like entertainment and sports
  • EbNERD contains more “hard news” like politics and sensitive topics

Filter Bubbles and Societal Risk

The interview spends significant time on how recommender systems can create or reinforce filter bubbles.

Mechanism

• Systems optimize for engagement.
• Engagement often means recommending more of what users already clicked.
• That creates a feedback loop:
  • user clicks a type of article
  • recommender keeps serving similar content
  • the user sees fewer opposing views

Observed effects

• Recommend systems may amplify already negative or polarizing framing.
• Even if the dataset is already biased, the recommender may make it worse.

Andrea’s view

• Diversity should be real diversity, not token balance.
• Conspiracy theories or fake news should not be promoted in the name of “balance.”

Balancing Accuracy, Diversity, and Stakeholders

A central argument of the episode is that news recommendation cannot be optimized by a single metric.

Why a single loss function is not enough

There are multiple, sometimes conflicting goals:

• relevance
• personalization
• viewpoint diversity
• category diversity
• multilingual access
• societal safety
• editorial standards

Multi-stakeholder perspective

Andrea argues for systems that consider:

• users
• editors
• writers
• platforms
• broader societal interests

A possible design

She suggests modular systems with adjustable knobs, where:

• relevance remains a baseline requirement
• users can request more diversity or more personalization
• system-level guardrails prevent harmful extremes

Where News Recommendations Appear in Real Life

Many users may not notice news personalization on a publisher’s own site, but it is common on aggregator platforms:

• Google News
• topic-based news apps
• personalized content feeds

These systems often infer interests from a few clicks and then keep surfacing similar stories or related local topics.

Future Directions for the Field

Andrea wants the field to move beyond “just making models bigger.”

Areas she wants more focus on

• reproducibility and fair comparisons
• diversity and serendipity
• multilingual and low-resource language support
• more realistic datasets
• multimodal news understanding
• responsible AI in information access

Broader vision

She wants systems that let users:

• access news in multiple languages
• avoid being trapped in a single perspective
• discover relevant, diverse content without needing to search across many platforms

Advice for People Entering the Field

Andrea recommends a multidisciplinary path.

Useful backgrounds

• machine learning
• natural language processing
• information retrieval
• social science
• media and communication
• psychology
• law

Why interdisciplinarity matters

Building the model is not enough; the field also needs people who understand:

• user behavior
• public discourse
• regulation
• fairness
• societal impact

What’s Next for Andrea

She is extending her work into adjacent areas:

Current interests

• Retrieval-augmented generation (RAG):
  • what happens when retrieved sources are made more diverse?
  • do generated answers reflect that diversity?
• Linguistic variation and dialects:
  • studying small-scale regional language differences
  • using dialects as a proxy for underrepresented spoken languages

Links and Online Presence

Andrea shared the following ways to follow her work:

• Personal website: andreiana.github.io
• Google Scholar: search for her name
• Social platforms:
  • LinkedIn
  • X
  • Bluesky

Main Takeaways

• News recommendation is fundamentally different from other recommendation problems because of timeliness, weak feedback, and societal impact.
• Performance alone is not a sufficient goal; diversity, inclusivity, and responsibility matter.
• In Andrea’s work, the news encoder matters more than the user encoder in many setups.
• Simpler models can match complex ones, especially when user feedback is sparse.
• Current benchmarks are limited by age, narrow modality, and lack of multilingual coverage.
• The future of news recommendation should be multimodal, multilingual, modular, and multi-stakeholder aware.