Build contextual quality signals
Created 41 indicators describing how news domains appear and circulate, then prepared reproducible datasets for modeling.
ContextEU-funded Social Media for Democracy research
My roleResearch Engineer responsible for data, modeling, recommendation, and evaluation
Verified result3-point source-quality gain with baseline precision maintained
Relevance alone is not enough.
Conventional recommenders optimize for what a user is likely to engage with. In a news environment, that objective can unintentionally promote unreliable domains. Product and research teams need a way to make information quality measurable without discarding personalization.
Add domain quality as an explicit ranking signal.
Validate against an independent benchmark
Compared learned domain-quality estimates with NewsGuard ratings to establish whether the signals captured meaningful reliability information.
Integrate quality into recommendation
Combined relevance-focused recommendation with learned domain-quality signals through a quality-aware re-ranking approach.
Evaluate the trade-off
Compared the new approach with the baseline using recommendation precision and average quality of the recommended domains.
Ownership across the research and engineering workflow.
Data and experimentation
Designed data preparation, feature engineering, validation checks, model comparison, and reproducible experiment workflows.
Recommendation design
Implemented and evaluated the recommendation pipeline, integrated learned signals, and communicated the relevance and quality trade-offs to collaborators.
Better source quality without lower baseline precision.
The quality-aware approach increased the average quality of recommended news domains by 3 points while maintaining the precision achieved by the relevance-focused baseline.
Where this capability creates value.
Media and content products
Balance engagement with trust, quality, safety, diversity, or editorial objectives that matter to the product.
Marketplaces and discovery systems
Add business-quality signals to ranking so relevance is not the only factor deciding what users see.
Responsible AI evaluation
Turn an abstract quality concern into signals, experiments, and measurable trade-offs that teams can inspect.
Recommender modernization
Evaluate an existing ranking pipeline and integrate additional objectives without replacing the complete product architecture.
Working on recommendation or ranking quality?
Share the current product, available data, and the trade-off you need to measure.