InductWave Cuts KG Query Layers by 75% While Outperforming Baselines
Wavelet-based embedding halves message-passing layers for inductive reasoning on massive KGs.
A team of researchers from IIIT-Delhi and Vrije Universiteit Amsterdam has proposed InductWave, a novel inductive embedding method for answering multi-hop logical queries on knowledge graphs (KGs). Current approaches rely on transductive reasoning, requiring all entities to be seen during training—a major limitation for real-world KGs that constantly grow. InductWave uses wavelet-based embeddings to generalize to unseen entities without retraining, achieving inductive reasoning with significantly less computational overhead.
On standard benchmarks like FB15k-237, InductWave matches or surpasses state-of-the-art models while using only half the number of message-passing layers—and in most cases, just 75% of the layers. This 25–50% reduction in layers directly translates to lower memory and compute requirements, enabling the model to scale to massive graphs like Wiki-KG. The team released code and datasets, positioning InductWave as a practical solution for resource-constrained environments where training on entire KGs is infeasible.
- InductWave reduces message-passing layers by 75% compared to baselines, cutting compute costs significantly.
- The wavelet-based method handles inductive reasoning—answering queries about entities never seen during training.
- Tested on FB15k-237 and Wiki-KG, it outperforms state-of-the-art models in most scenarios despite fewer layers.
Why It Matters
Enables scalable, inductive reasoning on massive knowledge graphs with half the compute, unlocking real-time query answering.