Dynamic multi-scale feature augmentation for inductive network representation learning

doi:10.1016/j.patcog.2024.111250

Dynamic multi-scale feature augmentation for inductive network representation learning

2025-05-01

0

PRE

AI

DOI：10.1016/j.patcog.2024.111250

Original

Original request for help

Save

Abstract

En 中文

Inductive Network Representation Learning (INRL) has been successfully applied in various graph-based machine learning tasks. Prior arts advance INRL by using Graph Neural Networks (GNNs) on graph-structured data. Most of them follow a topology-static aggregation scheme, where the representation of anode is calculated through a recursive process of aggregating and transforming the fixed local proximity information from its neighborhood. However, this may affect model generalization when facing data noise or data scarcity problems. Therefore, in this paper, we propose a novel INRL framework, dubbed MUFA, to learn meaningful network representations via dynamic MU lti-scale F eature A ugmentation based on GNNs. MUFA follows a topology-dynamic aggregation scheme, which incorporates structure-based and attribute-based graph properties as multi-scale features for data augmentation. Precisely, we design two modules to augment the features. One is the randomized combination of ego network structures, which can provide various substructures of graph data and relieve isolation issues caused by arbitrary selection of nodes. The other is the Inductive Masked Message Passing (IMMP), which dynamically masks parts of subregional information in the graph convolutional receptive field. Thus, diverse unmask-to-mask feature pairs are yielded as graph context augmentation for INRL. The pairs are fed to an auto-encoder, by which the encoder encodes the unmasked information and the decoder reconstructs networked relationships and reproduces node attributes in the masked subregions simultaneously. We conduct experiments on link prediction and node classification tasks over several public benchmarks. Experimental results present that the proposed MUFA performs better in model generalization and is able to generate high-quality network representations compared with state-of-the-art neural network baselines.

Keywords:

Inductive network representation learning

Graph neural networks

Multi-scale features

Data augmentation

Dynamic multi-scale feature augmentation for inductive network representation learning

Abstract

Journal

Organization