News recommender: retrieval, ranking, re-ranking

Problem

News decays fast. An article matters for hours, and most new articles have no clicks yet when they need to be recommended. The company's news platform needed recommendations that worked for that cold start and for sparse, long-tail content.

Approach

I designed and built an end-to-end recommender with a multi-stage architecture: candidate retrieval, deep learning ranking, and re-ranking for freshness and diversity.

• Retrieval. Two-tower retrieval with a Deep Structured Semantic Model (DSSM). The item tower encodes articles with a pretrained BERT-style language model, so newly published articles can be recommended before they collect clicks.
• Ranking. I benchmarked and deployed attention-based news rankers (NRMS, LSTUR). They model each user from their click history with multi-head self-attention over article representations.
• LLM layer. An LLM generates concise article and user-interest summaries. Those summaries feed richer semantic signals into retrieval and ranking for sparse and long-tail content.
• Cold start. Content-based and collaborative signals are blended, with recency-aware re-ranking on top.

Architecture

Interaction data comes from the data platform: GA4 events refined through a Bronze → Silver → Gold lakehouse into clean user-item interaction tables. Retrieval narrows the catalog to candidates, ranking scores them per user, and re-ranking balances freshness and diversity before the feed is served.

Results

The news and music recommenders together contributed to an approximately 40% increase in customer retention.