News Recommendation System

Delivered personalized news recommendations in a low-resource language environment.

Key Takeaways:

Language constraints limited recommendation quality

Limited NLP tooling made personalization difficult in low-resource languages.

We adapted transformer-based recommenders

A modern recommender architecture was applied and tuned for language constraints.

Transformer-based news recommendation model

An NRMS architecture with self-attention was implemented to model user preferences and article representations.

~30% gains in coverage and serendipity

Recommendation diversity improved while maintaining stable ranking performance.