KD-HGRL: Knowledge Distillation for Multi-Task Heterogeneous Graph Representation Learning

Nafiseh Imanian, Parham Moradi DW, laleh tafakori, Mahdi Jalili

Keywords: Knowledge Distillation, Graph Neural Network, Embedding Transfer, Heterogeneous Graph, Contrastive Learning, Self-Supervised Learning.

TL;DR: KDHGNN enhances representation learning in heterogeneous graphs through knowledge distillation and self-supervised contrastive learning for improved node classification and link prediction

Abstract: Heterogeneous graphs, characterized by diverse node and edge types, are central to many real-world applications, including social networks, biological systems, and recommendation engines. While Graph Neural Networks (GNNs) are effective for graph representation learning, their reliance on extensive labeled data, high computational cost, and long inference times limit scalability, especially for heterogeneous graphs. To address these challenges, we propose KD-HGRL, which leverages \textbf{K}nowledge \textbf{D}istillation for multi-task \textbf{H}eterogenous \textbf{G}raph \textbf{R}epresentation \textbf{L}earning. KD-HGRL uses self-supervised contrastive learning across semantic and topological views to generate robust, label-free node embeddings in the teacher phase. These embeddings are distilled into a lightweight student model, enabling efficient task-specific outputs such as node classification and link prediction with significantly reduced inference time. Experiments on benchmark datasets demonstrate KD-HGRL’s superior performance and efficiency compared to state-of-the-art methods. The framework captures both local and global graph structures, eliminates the need for labeled data, and scales effectively to large graphs. Key novelties, such as a multi-view teacher model, contrastive alignment, and a lightweight student model, make KD-HGRL a versatile and efficient solution for heterogeneous graph representation learning.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6380