Retrieval and Ranking Architectures

Retrieval-and-ranking architectures split recommendation into fast broad recall and slower precise ordering. This is the recommender analogue of dense retrieval followed by reranking: retrieve thousands, rank hundreds, display a few.

The two-stage system

A two-stage system can be written as

$$C_u={\mathrm{topK}}_{i\in \mathcal{I}}g(u,i),L_u={\mathrm{sort}}_{i\in C_u}h(u,i,x_{ui}),$$

where $g$ is a cheap retrieval score and $h$ is a richer ranking model. Candidate generation may combine multiple retrieval sources before ranking.

Worked example

A cheap retrieval score can select candidates before the ranker adds richer features:

Item	Retrieval score	Retrieved?	Ranker margin	Final rank score
0	0.74	yes	0.00	0.74
1	0.24	no	0.20	not ranked
2	0.60	yes	0.10	0.70
3	0.44	yes	0.30	0.74

The retrieval stage recalls items 0, 2, and 3; item 1 is never seen by the ranker. The ranker then promotes item 3 above item 2 using the margin feature. Hybrid recommenders often feed several such scores into the ranker.

Caveats

Retrieval and ranking must be evaluated separately. A ranker with excellent NDCG cannot fix low candidate recall, and a high-recall generator can still overload serving latency. Keep eligibility, freshness, deduplication, and exploration decisions visible in logs.

References

• Li et al., 2010, A Contextual-Bandit Approach to Personalized News Article Recommendation
• scikit-learn documentation: cosine_similarity

Nav

Section — Recommendation Systems and Personalization

← Ranking Evaluation of Recommenders →