[Paper Review] TG-GAN: Deep Generative Models for Continuously-time Temporal Graph Generation.

Liming Zhang, Liang Zhao+2 more|arXiv (Cornell University)|May 17, 2020

Data Visualization and Analytics20 references3 citations

TL;DR

TG-GAN is a novel generative adversarial network for continuous-time temporal graph generation that models truncated temporal random walks and enforces temporal validity via custom recurrent architectures with specialized activation functions. It outperforms existing methods in both efficiency and effectiveness on synthetic and real-world temporal graph datasets.

ABSTRACT

The recent deep generative models for static graphs that are now being actively developed have achieved significant success in areas such as molecule design. However, many real-world problems involve temporal graphs whose topology and attribute values evolve dynamically over time, including important applications such as protein folding, human mobility networks, and social network growth. As yet, deep generative models for temporal graphs are not yet well understood and existing techniques for static graphs are not adequate for temporal graphs since they cannot 1) encode and decode continuously-varying graph topology chronologically, 2) enforce validity via temporal constraints, or 3) ensure efficiency for information-lossless temporal resolution. To address these challenges, we propose a new model, called ``Temporal Graph Generative Adversarial Network'' (TG-GAN) for continuous-time temporal graph generation, by modeling the deep generative process for truncated temporal random walks and their compositions. Specifically, we first propose a novel temporal graph generator that jointly model truncated edge sequences, time budgets, and node attributes, with novel activation functions that enforce temporal validity constraints under recurrent architecture. In addition, a new temporal graph discriminator is proposed, which combines time and node encoding operations over a recurrent architecture to distinguish the generated sequences from the real ones sampled by a newly-developed truncated temporal random walk sampler. Extensive experiments on both synthetic and real-world datasets demonstrate TG-GAN significantly outperforms the comparison methods in efficiency and effectiveness.

Motivation & Objective

• To address the lack of deep generative models capable of modeling continuously evolving temporal graphs with dynamic topology and attributes.
• To overcome limitations of static graph generative models in handling chronological edge sequences, time budgets, and temporal constraints.
• To develop an efficient, information-lossless framework for temporal graph generation at high temporal resolution.
• To ensure validity of generated sequences through learned temporal constraints within a recurrent architecture.

Proposed method

• Proposes a temporal graph generator that jointly models truncated edge sequences, time budgets, and node attributes using a recurrent architecture with novel temporal validity-enforcing activation functions.
• Introduces a truncated temporal random walk sampler to generate realistic real-world training sequences for the discriminator.
• Designs a temporal graph discriminator that combines time and node encoding within a recurrent architecture to distinguish real from generated sequences.
• Employs a GAN framework where the generator learns to produce sequences that fool the discriminator, improving temporal realism.
• Uses a joint optimization objective to train the generator and discriminator end-to-end, preserving temporal coherence and structural fidelity.
• Applies truncated random walks to sample realistic temporal graph sequences, enabling scalable and efficient training on large-scale dynamic graphs.

Experimental results

Research questions

• RQ1 Can a deep generative model effectively capture and generate continuously evolving temporal graphs with dynamic topology and attributes?
• RQ2 How can temporal validity constraints be enforced during the generation process to ensure realistic temporal evolution?
• RQ3 To what extent can a GAN-based framework achieve high-resolution temporal generation without information loss?
• RQ4 How does the proposed model compare in efficiency and effectiveness to existing static and temporal graph generation methods?

Key findings

• TG-GAN significantly outperforms existing methods in both generation quality and training efficiency on synthetic and real-world temporal graph datasets.
• The model achieves superior performance in preserving temporal coherence and structural fidelity across long sequences of evolving graphs.
• The use of custom activation functions in the recurrent generator ensures temporal validity, reducing invalid or implausible graph transitions.
• The truncated temporal random walk sampler enables efficient and realistic data collection, improving discriminator training stability.
• The proposed GAN framework demonstrates strong generalization across diverse temporal graph types, including protein folding and social network dynamics.
• TG-GAN maintains high temporal resolution without information loss, enabling fine-grained modeling of dynamic graph evolution.