A hidden challenge of link prediction: which pairs to check?

The traditional setup of link prediction in networks assumes that a test set node pairs, which is usually balanced, available over to predict the presence links. However, practice, there no set: ground truth not known, so number possible pairs quadratic nodes graph. Moreover, because graphs are sparse, most these will be Thus, methods, often rely on proximity-preserving embeddings or heuristic notions similarity, face vast search space, with many close proximity, but should linked. To mitigate this issue, we introduce LinkWaldo, framework for choosing from quadratic, massively skewed space concise candidate that, addition being also structurally resemble observed edges. This allows it ignore some high-proximity low-resemblance and identify high-resemblance, lower-proximity pairs. Our built model theoretically combines stochastic block models (SBMs) proximity models. structure SBM maps out where new links expected fall, identifies plausible within blocks, using locality sensitive hashing avoid expensive exhaustive search. LinkWaldo can use any representation learning definition generate method, allowing power current future methods realized practice. We evaluate 13 across multiple domains show average returns sets containing 7–33% more missing than both embedding-based baselines’ sets. code at https://github.com/GemsLab/LinkWaldo.