The author manuscript, code, and reproduction materials for Invertible Graph Aggregation Can Increase ReLU Readout Width are now available on GitHub.
The study and public materials
Using a tree construction, the study shows that mixing neighboring features can preserve all information in the feature table while requiring more hidden units in a shallow classifier that reads each node separately. The classifier is a ReLU network with one hidden layer shared across all nodes. The number of hidden units in this layer is its width.
The author manuscript is version 1.0.1; code and reproduction materials remain version 1.0.0. The paper and reproduction instructions are available through the release and repository links below.
Paper and release details
- Paper
- Invertible Graph Aggregation Can Increase ReLU Readout Width
- Authors
- Hiroyasu Usami, Masanori Tsukamoto, Kohei Hashiguchi
- Manuscript
- Author preprint v1.0.1 (GitHub early release)
- Code and reproduction materials
- v1.0.0
- Submission
- Submitted to IEEE Access on 25 October 2026