Abstract
Learning low-dimensional representations for entities and relations inknowledge graphs using contrastive estimation represents a scalable andeffective method for inferring connectivity patterns. A crucial aspect ofcontrastive learning approaches is the choice of corruption distribution thatgenerates hard negative samples, which force the embedding model to learndiscriminative representations and find critical characteristics of observeddata. While earlier methods either employ too simple corruption distributions,i.e. uniform, yielding easy uninformative negatives or sophisticatedadversarial distributions with challenging optimization schemes, they do notexplicitly incorporate known graph structure resulting in suboptimal negatives.In this paper, we propose Structure Aware Negative Sampling (SANS), aninexpensive negative sampling strategy that utilizes the rich graph structureby selecting negative samples from a node's k-hop neighborhood. Empirically, wedemonstrate that SANS finds high-quality negatives that are highly competitivewith SOTA methods, and requires no additional parameters nor difficultadversarial optimization.