Unveiling Negative Sampling Strategies: A Comprehensive Guide for Recommender Systems

Recommender systems aim to capture personalized user preferences from massive interaction data, addressing information overload but suffering from data sparsity, dynamic user interests, filter bubbles, and feedback loops. Traditional models focus only on positive feedback, ignoring the crucial role of negative signals, which are often missing in datasets.

Negative sampling is essential for generating informative negative instances to balance training, yet it faces challenges such as false‑negative errors, trade‑offs between accuracy, efficiency, and stability, and generalization across tasks and datasets. This review fills the gap by systematically classifying and summarizing existing negative sampling research.

Existing negative sampling strategies are grouped into five major categories:

Static Negative Sampling

Dynamic Negative Sampling

Adversarial Negative Sample Generation

Importance Re‑weighting

Knowledge‑Enhanced Negative Sampling

Static Negative Sampling

Early deep recommender systems often rely on static negative sampling (SNS), selecting negatives from items a user has not interacted with. SNS aims to provide diverse negatives to capture a fuller user preference profile. Research is divided into four sub‑types: uniform, predefined, popularity‑based, and non‑sampling static strategies, each with distinct behaviors, benefits, and challenges.

Dynamic Negative Sampling

Dynamic Negative Sampling (DNS) selects informative negatives by evaluating candidate items against positive or user representations. It includes six groups: generic DNS, user‑similarity DNS, knowledge‑aware DNS, distribution‑based DNS, interpolation DNS, and hybrid DNS. Each approach balances deployment ease, reliance on user‑item scores, computational cost, and the ability to capture hard negatives.

Adversarial Negative Sample Generation

Adversarial Negative Sample Generation (ANG) enhances robustness by addressing the imbalance between abundant positives and scarce true negatives. Two paradigms exist: generative ANG, which uses GANs or other generative models to create high‑quality negatives, and sampling‑based ANG, which selects or re‑weights challenging negatives from existing candidate pools. Both improve discriminative ability but differ in computational demands and coverage of user preference complexity.

Importance Re‑weighting

Importance Re‑weighting (IRW) adjusts sample weights to emphasize more informative negatives. It includes attention‑based IRW, knowledge‑based IRW, and bias‑corrected IRW. Attention‑based methods allocate weights based on user interest signals, knowledge‑based methods leverage external structured knowledge for cold‑start scenarios, and bias‑corrected methods aim to mitigate systemic biases, balancing fairness and accuracy.

Knowledge‑Enhanced Negative Sampling

Knowledge‑Enhanced Negative Sampling (KNS) exploits auxiliary information such as user social contexts, heterogeneous item attributes, and knowledge graphs. It comprises generic KNS, which uses side information to select negatives closer to user preferences, and KG‑based KNS, which leverages entities and relations in a knowledge graph to uncover latent connections and choose more relevant negatives.

A comparative table (not reproduced here) summarizes representative methods across six classic recommendation models—collaborative filtering, graph‑based, sequential, multimodal, multi‑behavior, and cross‑domain—highlighting the negative sampling strategies each employs.

The review concludes with future research directions, including addressing false‑negative issues, curriculum learning for hard negatives, causal inference for understanding negative samples, and bias mitigation in sampling.