Recall Module in Recommendation Systems: Multi-Path Retrieval and Optimization

The article introduces the recall module of recommendation systems, which must handle hundreds of millions of items in the recommendation pool. Because a downstream ranking module provides a safety net, recall does not need to be perfectly accurate, but it must avoid missing relevant items and maintain low latency.

Recall is typically implemented via multi‑path (multi‑channel) retrieval, allowing parallel computation and complementary strengths. The two main categories are non‑personalized recall and personalized recall.

1. Significance of Recommendation Algorithms

With the rapid growth of users and content, recommendation systems have become essential for matching massive users with massive items. They complement search systems: search is user‑initiated, while recommendation is system‑initiated. The importance spans three perspectives:

For users: timely, accurate personalized content improves experience, activity, and retention.

For content producers: exposure incentives stimulate ecosystem growth.

For platforms: better recommendation drives higher DAU, conversion, ARPU, and overall business metrics.

2. Basic Modules of a Recommendation System

Recommendation Pool : Offline construction of a candidate pool (e.g., based on recent sales, video play counts).

Recall : Retrieves thousands to tens of thousands of items from the pool for the ranking stage. Must be lightweight and low‑latency; accuracy is secondary to coverage.

Coarse Ranking : A lightweight filter that reduces recall results to a few thousand items before fine ranking.

Fine Ranking : Precise scoring and ordering of candidates, often the most complex module.

Re‑ranking : Applies business rules, diversity, or contextual adjustments (e.g., holiday promotions).

Mixing : Merges results from multiple business lines (ads, videos, etc.) using rule‑based or RL methods.

3. Recall Details

3.1 Multi‑Path Recall

Recall faces an enormous candidate set, so multi‑path strategies are used to balance coverage and latency. The two major paths are:

Non‑personalized Recall : Offline‑built, user‑agnostic pipelines such as hot‑item recall, high‑efficiency recall, and operation‑strategy recall.

Personalized Recall : User‑specific pipelines, further divided into:

Content‑Based Recall

Tag recall (actors, directors, categories, etc.).

Knowledge‑graph based recall.

Multimodal recall (semantic similarity of titles, image similarity, video understanding).

Behavior‑Based Recall

User‑CF: Find users with similar behavior and use their interacted items.

Item‑CF: Find items similar to those the user has interacted with.

Feature‑Based Recall

Linear models (FM, FFM) using user/item features.

Deep models: twin‑tower DNNs (EBR, Mobius), deep match (YouTube‑DNN), sequence models (DIN, BERT4Rec), GNNs (Node2Vec, PinSage).

Online, recall can be performed via vector search (nearest‑neighbor on user embeddings) or i2i inverted indexes built from item embeddings.

3.2 Recall Optimization

Key challenges and solutions include:

Negative Sample Construction : Simple exposure‑without‑click samples are unsuitable. Hard negatives are generated using model scores, business rules, in‑batch negatives, or active learning.

Sample Selection Bias (SSB) : Random negatives may not represent the whole pool. Techniques such as hard‑negative mining, transfer learning (e.g., Alibaba ESAM), and balanced hard/easy negative ratios are employed.

Objective Misalignment : Recall focuses on similarity, while downstream ranking cares about conversion. Methods like CPM‑guided recall (Baidu Mobius) and joint model structures (Alibaba TDM) aim to align objectives.

Competition Among Paths : Overlap between recall channels reduces marginal gain. Effective merging and ensuring complementary coverage are necessary.

Overall, understanding the full recommendation architecture and the role of each module—especially recall—helps practitioners design efficient, scalable, and business‑aligned systems.