Abstract
Graph generation is the process of generating novel graphs with similar attributes to real world graphs. The explicit and precise control of granular structural attributes, such as node centrality and graph density, is crucial for effective graph generation. This paper introduces a controllable multi-objective translation model for text-attributed graphs, titled Controlled Graph Translator (CGT). It is designed to effectively and efficiently translate a given source graph to a target graph, while satisfying multiple desired graph attributes at granular level. Designed with an encoder-decoder architecture, CGT develops fusion and graph attribute predictor neural networks for controlled graph translation. We validate the effectiveness of CGT through extensive experiments on different genres of datasets. In addition, we illustrate the application of CGT in data augmentation and taxonomy creation, particularly in low resource settings.
Introduction
Graph generation—creating new graphs that mimic real-world networks—is vital for applications ranging from drug discovery to social network analysis. However, most existing models struggle with fine-grained control. Generating a molecule with a specific chemical stability or a citation network with a particular "closeness centrality" is difficult with current methods, which often generate graphs from scratch or focus on a limited set of properties. This research addresses the gap in "controlled transformation": the ability to take an existing source graph and precisely modify it to meet target structural criteria while maintaining its original textual content.
Methodology
The Controlled Graph Translator (CGT) utilizes an encoder-decoder architecture specifically tailored for graph-to-graph translation.
Core Components
- Dual Encoding: Standard Multi-Layer Perceptrons (MLPs) are used for encoding instead of Graph Neural Networks (GNNs). The authors found that GNNs often "oversmooth" representations, whereas MLPs better preserve the distinct features needed for accurate structural reconstruction. CGT encodes both the adjacency matrix (structure) and node features (textual content).
- The Control Vector: A dedicated Attribute Encoder processes source attributes, target attributes, and their difference to create a "control vector" that guides the structural modification.
- Element-wise Fusion: The graph representation and control vector are fused via element-wise multiplication, allowing the model to dynamically adjust feature influence during transformation.
- The Attribute Predictor: A pre-trained, frozen neural network estimates graph attributes (e.g., diameter, clustering coefficient) in real-time. This provides an "Attribute Loss" feedback loop during training to ensure the output matches desired structural properties.
Experimental Evaluation
The model was evaluated across diverse domains, including WordNet (semantic relationships), Ogbn-arxiv and Citeseer (citation networks), and MUTAG/MOLBACE (molecular biology). To train the translation capability, source-target pairs were created by randomly pruning edges from real-world graphs, challenging the model to reconstruct or modify the network to meet specific attribute targets.
Key Results
CGT consistently outperformed state-of-the-art baselines like GraphRNN, GT-GAN, and EDGE:
- Precision: On the WordNet dataset, CGT achieved a Mean Absolute Difference (MAD) of 0.56 in satisfying target attributes, a significant improvement over GraphRNN's 3.26.
- Data Augmentation: Using CGT to augment the MUTAG dataset led to a 10.5 point increase in accuracy for downstream property classification tasks.
- Semantic Understanding: In WordNet experiments, the model successfully "rediscovered" missing links, such as correctly connecting "ethnic group" to "egyptian," demonstrating an understanding of semantic context.
Discussion and Insights
A primary insight of this work is the effectiveness of framing graph generation as a translation task. By leveraging information from a source graph, the model avoids the difficulties of generating complex structures from scratch. Furthermore, the preference for MLPs over GNNs highlights a critical trade-off: while GNNs excel at neighborhood-based tasks, MLPs are more effective at preserving the granular structural details required for high-fidelity reconstruction. The model demonstrates true granular control, targeting metrics like Treewidth, Local Bridges, and Transitivity simultaneously.
Conclusion and Future Work
While CGT shows high precision in controlled transformation, scalability remains a challenge. Current tests involved graphs with up to 80 nodes; future iterations will need to explore sparse matrix representations for massive networks. Additionally, the authors plan to extend the framework to handle directed and weighted graphs and investigate meta-learning for "minimal training" scenarios in new domains.
Summary updated