时间：2026-09-07
本文（Structure-aware consistent graph autoencoder for generative content review）原载Expert Systems with Applications，由四川大学王竹教授等科研人员创作，系四川大学市场经济法治研究所系列学术成果。后续会持续分享四川大学市场经济法治研究所系列学术成果，欢迎大家阅读。
Traditional keyword and text classification methods primarily focus on local semantic features, limiting their ability to capture structural relationships inherent in generative content review. Graph neural networks (GNNs) provide a flexible paradigm to exploit graph topology for constructing the complex dependencies among targeted review objects, such as review elements, policy clauses and risk types. However, applying GNNs to link prediction-based generative content review still faces two key challenges: the mismatch of heterophilic links caused by homophily-oriented mechanisms and the potential instability of graph representation learning introduced by random augmentation strategies. This paper proposes a structure-aware consistent graph autoencoder (SCGAE) method for consistent topology predictions across different views in generative content review. The proposed method integrates the targeted review objects into a graph structure. It adaptively determines the graph type by measuring the distributional difference in feature similarity between positive and negative samples. This graph-type identification alleviates the topology mismatch caused by heterophilic links. A type-adaptive view augmentation strategy is employed to generate augmented views that preserve essential topological patterns. This strategy improves the stability of graph representation learning. Finally, SCGAE employs a cross-view consistency learning scheme to learn invariant representations across augmented views for topology reconstruction. Extensive experimental results on homophilic and heterophilic graph datasets demonstrate the effectiveness of the proposed SCGAE method.
Hanchen Zhang, Chuanbin Liu, Li Zhang, Zhu Wang, Zihao Yang, and Jie Chen. Structure-aware consistent graph autoencoder for generative content review, Expert Systems with Applications, vol. 333, 134261, Jan. 2027.（论文下载）
