Negative Sampling for Contrastive Representation Learning: A Review

TL;DR

This paper reviews negative sampling techniques in contrastive representation learning, highlighting their importance and categorizing existing methods to guide future research across various domains.

Contribution

It provides a systematic categorization of negative sampling methods in CRL and discusses their impact, filling a gap in the literature by focusing on negative sample selection.

Findings

Negative sampling methods are crucial for CRL success.

Four categories of negative sampling techniques are summarized.

Open research questions are identified for future work.

Abstract

The learn-to-compare paradigm of contrastive representation learning (CRL), which compares positive samples with negative ones for representation learning, has achieved great success in a wide range of domains, including natural language processing, computer vision, information retrieval and graph learning. While many research works focus on data augmentations, nonlinear transformations or other certain parts of CRL, the importance of negative sample selection is usually overlooked in literature. In this paper, we provide a systematic review of negative sampling (NS) techniques and discuss how they contribute to the success of CRL. As the core part of this paper, we summarize the existing NS methods into four categories with pros and cons in each genre, and further conclude with several open research questions as future directions. By generalizing and aligning the fundamental NS ideas across multiple domains, we hope this survey can accelerate cross-domain knowledge sharing and motivate future researches for better CRL.