Semi-Implicit Temporal Variational Graph Autoencoder for Dynamic Graph Generation

Graph simulation has emerged as a practical route to release realistic yet privacy-preserving network data, especially for temporal graphs that evolve through node and edge arrivals. However, prevailing approaches rely on temporal random walks that introduce sampling bias and become prohibitively costly at scale and impose Gaussian variational priors that fail to capture the heavy-tailed dynamics of real systems. We present a semi-implicit temporal variational graph autoencoder (SIT-VGAE) for high-fidelity temporal graph generation. SIT-VGAE samples localized ego-graphs and encodes them with a temporal graph attention network (TGAT) augmented with learnable time embeddings, capturing structural and temporal dependencies without random walk preprocessing. To overcome restrictive priors, we adopt semi-implicit variational inference in which a neural mixer defines an expressive, reparameterizable posterior family that better fits non-Gaussian dynamics. A lightweight decoder maps latent codes to timestamped edge distributions, whose assembled snapshots form generated temporal graphs. The model is trained to maximize a variational lower bound coupling a likelihood over edges with a Kullback–Leibler regularizer. Across multiple real temporal graph datasets, SIT-VGAE achieves superior simulation fidelity on structural statistics, while reducing training and generation time compared with random-walk-based and Gaussian prior baselines. The resulting framework scales to large graphs via TGAT encoding and amortized inference, offering a practical, efficient, and expressive solution for realistic temporal network synthesis.