[Paper Review] GraphTER: Unsupervised Learning of Graph Transformation Equivariant Representations via Auto-Encoding Node-wise Transformations

Xiang Gao, Wei Hu+1 more|arXiv (Cornell University)|Nov 19, 2019

3D Shape Modeling and Analysis54 references4 citations

TL;DR

This paper proposes GraphTER, an unsupervised method for learning graph representation by auto-encoding node-wise transformations—applying isotropic or anisotropic transformations to randomly sampled node groups. By reconstructing these transformations from encoded features, GraphTER captures local and global structural invariants, achieving state-of-the-art performance on 3D point cloud segmentation (81.9% mIoU), significantly outperforming prior unsupervised methods and nearing fully supervised performance.

ABSTRACT

Recent advances in Graph Convolutional Neural Networks (GCNNs) have shown their efficiency for non-Euclidean data on graphs, which often require a large amount of labeled data with high cost. It it thus critical to learn graph feature representations in an unsupervised manner in practice. To this end, we propose a novel unsupervised learning of Graph Transformation Equivariant Representations (GraphTER), aiming to capture intrinsic patterns of graph structure under both global and local transformations. Specifically, we allow to sample different groups of nodes from a graph and then transform them node-wise isotropically or anisotropically. Then, we self-train a representation encoder to capture the graph structures by reconstructing these node-wise transformations from the feature representations of the original and transformed graphs. In experiments, we apply the learned GraphTER to graphs of 3D point cloud data, and results on point cloud segmentation/classification show that GraphTER significantly outperforms state-of-the-art unsupervised approaches and pushes greatly closer towards the upper bound set by the fully supervised counterparts. The code is available at: https://github.com/gyshgx868/graph-ter.

Motivation & Objective

• To address the high labeling cost of supervised graph representation learning in irregular data such as 3D point clouds.
• To learn graph representations that are invariant to both global and local structural transformations without requiring labeled data.
• To develop a self-supervised framework that reconstructs node-wise transformations to enforce structural awareness in learned features.
• To enable unsupervised pre-training that generalizes well to downstream node-level tasks like segmentation and classification.
• To bridge the performance gap between unsupervised and fully supervised graph representation learning on 3D geometric data.

Proposed method

• The method samples random groups of nodes globally or locally and applies node-wise isotropic or anisotropic transformations (e.g., translation, rotation, shearing) to each group.
• An encoder network learns joint representations of the original graph and its transformed version, using EdgeConv blocks with skip connections to preserve local structure.
• A transformation decoder reconstructs the applied node-wise transformations from the encoded features of both original and transformed graphs.
• The model is trained end-to-end via a contrastive reconstruction loss that encourages the decoder to recover the transformation applied to each node group.
• The architecture uses a shared encoder for both original and transformed graphs, with a shared decoder head for transformation reconstruction.
• During training, the first four EdgeConv blocks are frozen to preserve structural feature learning, and the final classifier is applied node-wise on concatenated features.

Experimental results

Research questions

• RQ1 Can unsupervised graph representation learning be improved by explicitly modeling node-wise transformations as supervisory signals?
• RQ2 How do isotropic versus anisotropic node-wise transformations affect the quality of learned graph representations?
• RQ3 Does local sampling of nodes for transformation application yield better structural awareness than global sampling?
• RQ4 To what extent can an auto-encoding framework for node-wise transformations close the performance gap between unsupervised and fully supervised graph learning on 3D point clouds?
• RQ5 Can transformation-equivariant representations learned via this method generalize to downstream tasks like part segmentation and classification?

Key findings

• GraphTER achieves a mean Intersection-over-Union (mIoU) of 81.9% on the ShapeNet part segmentation benchmark, significantly outperforming the prior state-of-the-art unsupervised method MAP-VAE by 13.9 percentage points.
• The model performs comparably to fully supervised state-of-the-art methods such as DGCNN and RS-CNN, demonstrating that unsupervised pre-training can approach fully supervised performance.
• Shearing transformations yield the highest mIoU (81.13%), outperforming translation (79.90%) and rotation (80.35%) on average, indicating that complex transformations enhance structural learning.
• Local sampling of nodes for transformation application improves performance by 0.14% over global sampling (80.53% vs. 80.39%), highlighting the importance of local structural awareness.
• Qualitative results show that GraphTER produces more accurate segmentation, especially on fine-grained parts like plane engines and chair legs, compared to MAP-VAE.
• Ablation studies confirm that the proposed method is robust across different transformation types and sampling strategies, with consistent gains in representation quality.