Recall Stage in Recommendation Systems: From Intuition to Deep Learning

Recommendation systems aim to automatically select the most relevant items for users, saving time and creating value. The first step after candidate generation is the recall stage, which quickly narrows down millions of items to a manageable set for later ranking.

Recall is often described as match (similar to a first‑date screening) or candidate generation (similar to shortlisting resumes). Simple attribute‑based recall (e.g., sorting by sales, freshness, click‑through rate) is intuitive but suffers from low coverage and lack of personalization.

To overcome these limits, matrix factorization was introduced. By constructing a user‑item interaction matrix and factorizing it into low‑dimensional latent vectors, we can compute similarity between users, items, and user‑item pairs. This leads to collaborative‑filtering methods such as item‑CF, user‑CF, and model‑based CF.

Embedding techniques further enrich vectors with side information (user demographics, item categories, etc.) and enable deep learning models. The popular dual‑tower architecture processes users and items separately, produces comparable embeddings, and retrieves the nearest‑neighbor items in real time (often within 10 ms for millions of items).

Recent advances include tree‑based models like Alibaba’s TDM and ByteDance’s Deep Retrieval, which address diversity and complex user‑item interactions beyond the basic dual‑tower approach.

Overall, recall remains a critical, evolving component of recommendation pipelines, balancing efficiency, diversity, personalization, and real‑time constraints.