.png){kind=icon}
.png){kind=icon}
Return
H-HRGAN: Knowledge Graph-Driven Representation for Missing Value Imputation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.130067.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• H-HRGAN Framework: We introduce the Heterogeneous-Homogeneous Relational Graph Attention Network (H-HRGAN), a new framework designed for imputing missing values in structured data. By representing discrete feature values as nodes and attributes as relations, H-HRGAN effectively captures complex semantic relationships and inter-feature dependencies, overcoming the limitations of traditional imputation methods. • Relational Subgraph Attention Mechanism: We present a relational subgraph learning mechanism that enhances context-aware imputation by capturing detailed dependencies across structured data elements, increasing the accuracy of missing value predictions. • Multi-Scale Dynamic Convolution with Channel Attention: We utilize multi-scale dynamic convolution to extract vector information through Graph Neural Networks (GNNs), with a channel attention mechanism to assign different weights to various convolutional scales, further improving the modeling of entity-relation interactions. • Experimental Validation: Extensive experiments on diverse real-world datasets shows that H-HRGAN consistently outperforms state-of-the-art imputation methods under both missing-at-random (MAR) and missing-not-at-random (MNAR) conditions, achieving superior accuracy and robustness, especially for structured data with complex dependencies.
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W