InductWave trains knowledge graph models with half as many layers as competing approaches

A team of researchers has introduced InductWave, a wavelet-based inductive embedding method for answering multi-hop logical queries over knowledge graphs. It uses half as many message-passing layers as baseline models. The paper was published on arXiv on July 7, 2026.

The problem with transductive models

Multi-hop logical queries over knowledge graphs (Multi-Hop Query Answering) are formulated as queries that implicitly assume the data is complete, and they typically include conjunction, disjunction, and negation operators (Existential First Order Logic). Most existing models rely on transductive reasoning, which means they cannot reason about entities they did not see during training. In practice, that is a serious limitation: training a model on every node of a large knowledge graph at once is often impossible because of limited resources.

How InductWave works

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

• The method is built on wavelet embeddings
• With half the number of message-passing layers, InductWave performs on par with baseline models
• With 75% of the baseline models' layer count, InductWave outperforms them in most cases
• Lower resource requirements made it possible to test the model on massive graphs, including Wiki-KG
• The experiments covered different ratios of training to test graphs in the FB15k-237 dataset

The model's code and datasets are publicly available on GitHub.

What it means

Fewer message-passing layers directly cut the compute cost of training and inference. As a result, inductive methods like InductWave make it practical to work with truly large knowledge graphs such as Wiki-KG without massive compute clusters.

FAQ

What is inductive learning on knowledge graphs?

It is an approach in which a model trains 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.

Which datasets was InductWave tested on?

The model was evaluated on different ratios of training to test graphs in the FB15k-237 dataset, as well as on the large Wiki-KG knowledge graph, where its lower resource requirements proved especially important.