Skip to content
Conference Open access

Graph Autoencoders for Synthetic Tree Skeletons: A Sparse Topology Reconstruction Baseline

Oct 2026 · International Conference on Information and Software Technologies · pp. 537-548 · 0 citations · 21 references

Abstract

Tree skeletons are naturally represented as graphs, in which nodes describe biological or geometric elements and edges represent parent–child branch connections. This work presents a baseline study of graph autoencoders for latent representation learning and connectivity reconstruction of synthetic tree skeletons generated using an EvoEngine/EcoSysLabProject workflow based on the Rhizomorph procedural growth model. The dataset contains 16 synthetic tree instances from four species, with four individuals per species. Each tree snapshot is represented as a graph containing node attributes, spatial positions, and parent–child connectivity.A graph autoencoder with a three-layer graph convolutional encoder and an inner-product decoder is trained to reconstruct connectivity from node attributes and spatial positions. Three normalization strategies are compared: input normalization, latent normalization, and their combination. The results show that input normalization provides the most effective training behavior among the tested configurations and is therefore used for the subsequent reconstruction analysis.Full-adjacency evaluation shows that the model assigns high scores to nearly all ground-truth edges but also to a much larger number of non-edges. Consequently, threshold-based decoding produces graphs that are substantially denser than the ground-truth tree skeletons, with near-complete edge recall but negligible precision. This failure cannot be explained solely by the choice of decision threshold: when the number of predicted edges is constrained to match the number of ground-truth edges, only 8.03% of the true directed edges are recovered. Ranking-based evaluation further indicates that the decoder scores contain some information about connectivity, but do not rank the true edges sufficiently well for sparse topology reconstruction.These results show that a standard graph autoencoder trained with sampled negative edges is insufficient for recovering the connectivity of the evaluated tree skeletons. The main contribution of this work is therefore a baseline evaluation and failure analysis that identifies limitations of sampled-negative training and unconstrained inner-product decoding for sparse tree graphs. The analysis motivates future work on hard-negative mining, imbalance-aware objectives, geometry-aware decoders, topology-constrained graph extraction, and parent-prediction formulations.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.