InductWave Improves Inductive Logical Query Answering on Large Knowledge Graphs.

Logical Multi-Hop Query Answering (LMQ) on Knowledge Graphs (KGs) is a critical task, often assuming completeness and primarily focusing on Existential First Order Logic (EFO) queries. A major limitation of current methods is their transductive nature, meaning they cannot reason about entities not present during training. This is problematic in real-world scenarios where KGs are vast, and it's impractical to train models with all nodes. To address this, the researchers propose InductWave, a novel wavelet-based inductive embedding method specifically designed for LMQ on large KGs. InductWave allows for reasoning over entities that were not seen during the training phase, which is a significant advantage for dynamic and expanding knowledge bases. The model demonstrates strong performance, matching or exceeding baseline models while requiring significantly fewer message-passing layers—half the number in some cases, and outperforming them in most cases with 75% of the layers. This reduction in resource requirements enables InductWave to be evaluated on massive graphs, such as Wiki-KG. Extensive experiments on varying train-test graph proportions of the FB15k-(237) dataset confirm its superiority over state-of-the-art models.