Transductive Dynamic Graph Generation

• Transductive Dynamic Graph Generation (TDGG) is a framework for forecasting dynamic graph snapshots with fixed node sets and evolving edge structures and text attributes.
• It employs hybrid LLM and GNN strategies through multi-agent pipelines, reinforcement learning, and intensity-free modeling to simulate social interactions and textual content.
• TDGG's evaluation leverages structural, textual, and discriminative metrics, demonstrating scalability, memory integration, and precise interaction alignment.

Transductive Dynamic Graph Generation (TDGG) is a rigorous generative framework for modeling and forecasting dynamic graphs where the set of nodes is fixed, but the network's temporal structure and rich textual attributes on both nodes and edges evolve over time. TDGG is central to contemporary research on dynamic text-attributed graphs (DyTAGs), LLM-driven simulation, and scalable generative modeling of temporal graphs (Ji et al., 28 Oct 2025, Peng et al., 4 Jul 2025, Gupta et al., 2022).

1. Formal Definition and Mathematical Framework

TDGG is defined by the task of generating a future dynamic graph snapshot—including edge structure, temporal details, and associated rich text attributes—given a fixed node set and a temporal sequence of prior graph snapshots. The standard DyTAG formalism for TDGG, as instantiated in GDGB, is

$$G^{(t)} = (\mathcal{N}, \mathcal{E}^{(t)}, \mathcal{T}^{(t)}, \mathcal{N}^\text{text}, \mathcal{E}^\text{text}, \mathcal{L})$$

where:

• $\mathcal{N}$: set of nodes (fixed, known a priori)
• $$\mathcal{E}^{(t)} \subseteq \mathcal{N} \times \mathcal{N}$$: edges at time $t$
• $\mathcal{T}^{(t)}_{u,v}$: timestamp for edge $(u,v)$
• $\mathcal{N}^\text{text}$: node-level raw text attributes
• $\mathcal{E}^\text{text}$: edge-level raw text (e.g., messages, reviews)
• $\mathcal{L}$: categorical edge labels

Given $\{G^{(1)}, ..., G^{(K-1)}\}$, TDGG seeks to generate

$\mathcal{N}$0

maximizing the conditional probability $\mathcal{N}$1, given model parameters $\mathcal{N}$2 (Peng et al., 4 Jul 2025).

In social simulation contexts, the generation process is factored as

$\mathcal{N}$3

where the active source nodes $\mathcal{N}$4 are given for each $\mathcal{N}$5, and the core generative challenges are destination selection and edge generation (Ji et al., 28 Oct 2025).

2. Model Architectures and Generation Pipelines

Numerous frameworks formalize and operationalize TDGG, unifying traditional dynamic GNNs, LLM-based agents, and temporal generative models.

Agent-Driven LLM Approaches

• Graphia employs two Qwen3-8B-based agents:
  • Graphia-Q (for destination selection): Input is the node's memory and profile; output is a query plus behavioral filter, followed by BERT-embedded neighbor retrieval and rule-based filtering.
  • Graphia-E (for edge generation): Input includes source/target profiles and their interaction history; target is to generate ground-truth message and edge category.
  • Both agents first undergo supervised fine-tuning, then reinforcement learning (GRPO) with hybrid GNN-derived and text-based rewards (Ji et al., 28 Oct 2025).
• GAG-General (GDGB benchmark) implements TDGG as a multi-agent pipeline, with each node as an autonomous LLM agent holding memory of past interactions. The generation loop alternates between destination recall/selection and edge text/label generation, updating agent memories for each round. LLM-based "memory reflection" and history-augmented prompts improve semantic consistency and structure alignment (Peng et al., 4 Jul 2025).

Temporal Point Process and Random Walk Models

• TIGGER parameterizes TDGG as an intensity-free temporal point process over the fixed node set. The generation decomposes into auto-regressive random walks, with next-node and next-time sampled from mixture log-normal models conditioned on node ID embeddings and LSTM history. Edges are generated iteratively, directly leveraging the original node identities (hence fully transductive) (Gupta et al., 2022).

3. Evaluation Metrics and Benchmarking Protocols

TDGG requires multifaceted evaluation to assess both structural fidelity and semantic/temporal alignment.

Metric Category	Example Metrics	Description
Structural	Degree/Spectra MMD, Power-law	Maximum Mean Discrepancy of degree and spectra; power-law exponent and validity; graph-level statistics (Peng et al., 4 Jul 2025, Gupta et al., 2022)
Textual	LLM-as-Judge Scores, ROUGE-L, BERTScore	Human-aligned evaluation (contextual fidelity, personality depth, adaptability, immersive quality, content richness) and classical text metrics (Ji et al., 28 Oct 2025, Peng et al., 4 Jul 2025)
Micro-level interaction	R@100 (Easy/Hard/All), Acc, S_TDGG	Recall-at-100 for destination, category accuracy, node-wise composite scores (Ji et al., 28 Oct 2025)
Embedding	Extended JL-Metric	Frobenius-norm cosine distance between low-dim node embeddings (text, structure, time) (Peng et al., 4 Jul 2025)
Discriminative	Link pred., Edge Cla., F1	Hit@K, F1 on link/edge label classification, typically for analysis not training (Peng et al., 4 Jul 2025)

Normalization (min-max, per-dataset) ensures comparability across datasets and conditions.

4. Datasets and Experimental Settings

GDGB establishes a standardized benchmark of eight multi-domain DyTAGs with high textual fidelity. Key properties:

• Bipartite and non-bipartite graphs (e.g., Sephora, Dianping, WikiRevision, IMDB, WeiboTech, WeiboDaily, Cora)
• Rich node and edge texts, ranging from short bios to long review texts
• Time-resolved snapshots with diverse time granularities (from milliseconds to months)
• Large scale: node counts $\mathcal{N}$6k--$\mathcal{N}$7k, edges $\mathcal{N}$8k--$\mathcal{N}$9M, multiple categories/labels (Peng et al., 4 Jul 2025)

Experimentally, all frameworks initialize with seed graphs and predict multi-round future expansions (e.g., 50 edges per round, 10,000 total for TDGG) (Peng et al., 4 Jul 2025). Social simulation datasets in Graphia include Propagate-En, Weibo Daily, and Weibo Tech (Ji et al., 28 Oct 2025).

5. Empirical Findings and Benchmark Results

Several non-trivial findings characterize TDGG's empirical landscape:

• Structural Fidelity: GAG-General (LLM agents) achieves low degree MMD (0.02–0.28) and power-law validity on 6/8 GDGB datasets, rivaling or surpassing traditional DGNNs (Peng et al., 4 Jul 2025). In Graphia, hybrid RL agents yield a 41.11% improvement in macro-level structural similarity versus the best baseline (Ji et al., 28 Oct 2025).
• Textual Quality: Average LLM-as-judge ratings consistently exceed 4 on a 5-point scale with memory and reflection mechanisms (Peng et al., 4 Jul 2025). Graphia's BERTScore-F1 achieves a 27.9% absolute lift over the strongest competitors (Ji et al., 28 Oct 2025).
• Interaction Alignment: Composite node-wise scores (S_TDGG) show best-in-class micro-level fidelity: e.g., 0.937 on Propagate-En (vs. baseline 0.639), 0.965 on Weibo Tech, and a 12% absolute gain in edge category accuracy (Ji et al., 28 Oct 2025).
• Discriminative Task Transfer: Generative TDGG agents (GAG-General) often outperform DGNNs trained on sparse edges in link prediction and edge classification (e.g., F1 up to 0.79 vs. 0.53 on Sephora/IMDB) (Peng et al., 4 Jul 2025).
• Scalability: TIGGER generates transductive dynamic graphs up to $$\mathcal{E}^{(t)} \subseteq \mathcal{N} \times \mathcal{N}$$0 nodes and edges with $$\mathcal{E}^{(t)} \subseteq \mathcal{N} \times \mathcal{N}$$1–$$\mathcal{E}^{(t)} \subseteq \mathcal{N} \times \mathcal{N}$$2 speedup compared to prior art, maintaining low duplication and median errors on 8/10 structural metrics (Gupta et al., 2022).

6. Key Algorithmic Insights and Theoretical Principles

• Memory Integration: Node-wise memory of past interactions (and periodic LLM reflections) is critical for aligning generated textual content and structural evolution, as ablations show substantial quality drops without memory (Peng et al., 4 Jul 2025).
• Structural Rewarding: Incorporation of GNN-derived feedback (e.g., GraphMixer) into RL reward functions directly improves edge label and category accuracy and closes the gap between LLM predictions and true network dynamics (Ji et al., 28 Oct 2025).
• Hybrid Objective Functions: Joint optimization for both destination selection and rich edge text/content is necessary to avoid myopic, non-representative generation. Ablations that ignore higher-order graph structure or use only sequential data produce no better outcomes than mere SFT (Ji et al., 28 Oct 2025).
• Agent Collaboration: Multi-agent paradigms (LLM agents for each node) facilitate holistic simulation and ensure both local (micro-level) and emergent global (macro-level) properties emerge in generated graphs (Peng et al., 4 Jul 2025).
• Intensity-Free Modeling: TIGGER demonstrates that mixture density-based modeling of inter-event times, rather than direct intensity parameterization, allows intensity-free log-likelihood maximization and avoids computationally expensive point process training (Gupta et al., 2022).

7. Applications, Scope, and Open Challenges

TDGG provides a fundamental testbed for:

• LLM-based social simulation with realistic agent behaviors and network phenomena (e.g., echo chambers, power-law degree) (Ji et al., 28 Oct 2025)
• Scenario counterfactuals, e.g., simulating policy or incentive shifts on social platforms
• Generative modeling for dynamic recommender systems, user-to-item graphs, and knowledge graphs with evolving edge semantics (Peng et al., 4 Jul 2025)
• Fast, large-scale generation and benchmarking of dynamic graphs for downstream discriminative and generative tasks (e.g., link prediction, node anomaly detection) (Gupta et al., 2022)

Persistent open challenges include capturing global long-range dependencies, handling extreme text lengths and out-of-distribution semantics, and bridging the micro-macro gap for emergent network properties under non-i.i.d. conditions.

• "GRAPHIA: Harnessing Social Graph Data to Enhance LLM-Based Social Simulation" (Ji et al., 28 Oct 2025)
• "GDGB: A Benchmark for Generative Dynamic Text-Attributed Graph Learning" (Peng et al., 4 Jul 2025)
• "TIGGER: Scalable Generative Modelling for Temporal Interaction Graphs" (Gupta et al., 2022)