InductWave Trains Knowledge Graph Models with Half the Layers of Competitors

A team of researchers presented InductWave—a method of inductive embedding based on wavelets for answering multi-hop logical queries to knowledge graphs, which works with twice as few message-passing layers as baseline models. The paper was published on arXiv on July 7, 2026.

What's the problem with transductive models

Multi-hop logical queries to knowledge graphs (Multi-Hop Query Answering) are formulated as queries with an implicit assumption of data completeness and typically contain operators of conjunction, disjunction, and negation (Existential First Order Logic). Most existing models use transductive reasoning—meaning they cannot reason about entities they did not see during training. In the real world, this is a serious limitation: training a model on all nodes of a large knowledge graph at once is often impossible due to resource constraints.

How InductWave is structured

InductWave is an inductive method: the graph on which the model is trained contains fewer nodes than the graph on which it is tested.

• The method is based on wavelet embeddings
• With half the number of message-passing layers, InductWave shows results on par with baseline models
• With 75% of the number of layers of baseline models, InductWave exceeds them in most cases
• More modest resource requirements allowed the model to be tested on massive graphs, including Wiki-KG
• Experiments were conducted on different proportions of training and test graphs from the FB15k-237 dataset

The model's code and datasets have been made publicly available on GitHub.

What this means

Fewer message-passing layers directly reduces computational costs for both training and inference—meaning inductive methods like InductWave make working with truly large knowledge graphs, such as Wiki-KG, practically feasible without enormous computing clusters.

Frequently asked questions

What is inductive learning on knowledge graphs?

This is an approach where a model is trained on a graph with fewer nodes than the one it is later tested on, and must be able to reason about entities it did not see during training—unlike transductive models.

What datasets was InductWave tested on?

The model was tested on different proportions of training and test graphs from the FB15k-237 dataset, as well as on the large knowledge graph Wiki-KG, where lower resource requirements proved particularly important.