摘要
Analyzing attribute-missing graphs with a complete topology, but missing the attributes of some nodes, is an emerging and challenging research topic. Data imputation techniques based on graph autoencoders are commonly used for attribute-missing graphs. However, this method cannot effectively integrate existing attributes and structural information during the encoding stage and is prone to introducing noise, resulting in inaccurate imputation. In addition, the expressiveness of decoders in existing methods is limited because their network architecture has not been adequately designed, which restricts the accuracy and robustness of the generated attributes. To address these issues, we propose a novel Attribute Imputation AutoEncoder for attribute-missing graphs, named AIAE. In particular, during the encoding stage, a dual encoder based on knowledge distillation is designed to encode both attribute and structural information into representations of attribute-missing nodes to achieve more accurate imputation. To avoid introducing noise, we fully exploit the observed information by reorganizing the representations of the attribute-missing and attribute-observed nodes. In the decoding stage, we propose a multi-scale decoder with masking to make the decoder more expressive and enhance its robustness and generative ability. Extensive experiments demonstrate that our model significantly outperforms state-of-the-art methods in attribute-missing graphs.
摘要译文
通过完整的拓扑分析属性错失图形，但缺少某些节点的属性是一个新兴而具有挑战性的研究主题。基于图形自动编码器的数据插补技术通常用于属性失误图。但是，此方法无法在编码阶段有效地整合现有属性和结构信息，并且容易引入噪声，从而导致淘汰不准确。此外，现有方法中解码器的表现力受到限制，因为它们的网络体系结构尚未经过充分的设计，这限制了生成属性的准确性和鲁棒性。为了解决这些问题，我们提出了一种新颖的属性自动编码器，用于属性错失图，名为AIAE。特别是，在编码阶段，基于知识蒸馏的双重编码器旨在将属性和结构信息编码为属性损坏节点的表示形式，以实现更准确的插补。为了避免引入噪声，我们通过重新组织属性失误和属性观察的节点的表示来充分利用观察到的信息。在解码阶段，我们提出了一个带有掩盖的多尺度解码器，以使解码器更具表现力并增强其鲁棒性和生成能力。广泛的实验表明，我们的模型在属性错失图中的最先进方法显着胜过。
Bo Yang [a] [c]. Attribute imputation autoencoders for attribute-missing graphs[J]. Knowledge-Based Systems, 2024,291: 111583