Similarity-Constrained Reweighting for Complex Query Answering on Knowledge Graphs

ICLR 2026 Conference Submission4382 Authors

Keywords: knowledge graphs, query answering, similarity

Abstract:

Machine learning models for answering complex queries on knowledge graphs estimate the likelihood of answers that are not reachable via direct traversal. Prior work in this area has focused on structured queries whose constraints are expressed in first-order logic. Recent work has proposed to extend such logical constraints with *soft entity constraints*, which require the answer to a query (also known as the *target variable*) to be similar or dissimilar to specified sets of entities.

A natural but unexplored generalization of this extension is to allow specifying similarity constraints not only on the answer to a query, but also on the values assigned to *intermediate variables*, which frequently occur in complex queries. In this work, we study this more general formulation and introduce SCORE: a computationally efficient and interpretable method for incorporating similarity constraints at arbitrary positions of a query. Unlike approaches that rely on deep neural networks, SCORE is based on a lightweight and interpretable score adjustment function that only requires tuning two parameters on a validation set.

Our experiments on a challenging benchmark over three different knowledge graphs demonstrate that on the special case of constraints on the target variable, SCORE is able to incorporate preferences without degrading overall query answering performance, with significantly increased ranking performance over a learned neural baseline. Moreover, SCORE maintains its performance in the more general setting with constraints on intermediate variables.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 4382