مدل مولد گرافي عميق سلسله مراتبي و آگاه از جامعه

Generating realistic graphs an‎d efficiently sampling from their probability distributions is considered one of the most challenging problems in deep generative learning, due to the high dimensionality of the space an‎d the complex dependencies among edges. Traditional models, relying on simple, predefined structural assumptions, fail to adequately capture the intricate patterns of the real world. In contrast, existing deep models also face significant limitations: one-shot approaches, assuming edge independence in heterogeneous graphs, yield low-quality results, while autoregressive approaches, though better at modeling dependencies, encounter severe scalability issues in large-scale graphs an‎d suffer a significant dro‎p in output quality when confronted with communities exhibiting diverse structural distributions. In this dissertation, a community-aware hierarchical framework named HiCGG is proposed, which simultaneously addresses three key challenges: modeling structural heterogeneity, computational scalability, an‎d flexibility in generating hybrid topologies. Leveraging a divide-an‎d-conquer strategy, this framework recursively partitions the graph into homogeneous communities, independently generates each community using deep autoregressive generative models, an‎d then accurately reconstructs inter-community edges through an aggregation module; an advanced end-to-end version is also introduced, integrating deep learning-based community detection. In extensive eva‎luations on both synthetic an‎d real-world datasets, the HiCGG model significantly outperformed state-of-the-art methods: in heterogeneous graphs, it achieved an average improvement of 65% in classical statistical metrics an‎d 30–40% in graph neural network-based metrics, while in large-scale graphs it delivered the highest quality while maintaining a substantial advantage across all metrics. The model reduced graph generation time by up to 75% with increasing hierarchy depth an‎d, thanks to its modular, scalable, an‎d community-aware architecture, overcame fundamental limitations of prior models, opening new horizons in fields such as drug discovery, protein modeling, an‎d fully realistic synthetic data generation

عبدالرضا ميرزايي

مهران صفاياني

زينب مالكي , محمدرضا احمدزاده , عماد فاطمي زاده