Beyond Smoothing: Unsupervised Graph Representation Learning with Edge Heterophily Discriminating

Abstract

Unsupervised graph representation learning (UGRL) has drawn increasing research attention and achieved promising results in several graph analytic tasks. Relying on the homophily assumption, existing UGRL methods tend to smooth the learned node representations along all edges, ignoring the existence of heterophilic edges that connect nodes with distinct attributes. As a result, current methods are hard to generalize to heterophilic graphs where dissimilar nodes are widely connected, and also vulnerable to adversarial attacks. To address this issue, we propose a novel unsupervised Graph Representation learning method with Edge hEterophily discriminaTing (GREET) which learns representations by discriminating and leveraging homophilic edges and heterophilic edges. To distinguish two types of edges, we build an edge discriminator that infers edge homophily/heterophily from feature and structure information. We train the edge discriminator in an unsupervised way through minimizing the crafted pivot-anchored ranking loss, with randomly sampled node pairs acting as pivots. Node representations are learned through contrasting the dual-channel encodings obtained from the discriminated homophilic and heterophilic edges. With an effective interplaying scheme, edge discriminating and representation learning can mutually boost each other during the training phase. We conducted extensive experiments on 14 benchmark datasets and multiple learning scenarios to demonstrate the superiority of GREET.

Refining Latent Homophilic Structures over Heterophilic Graphs for Robust Graph Convolution Networks

“…Recently we have witnessed a surge in the robustness of GCN on heterophilic graphs. These methods can be categorized into the structure learning-based ones (Jin et al 2020(Jin et al , 2021aHe et al 2022;Zhu et al 2022;Liu et al 2023), and the ones based on adversarial training (Dai et al 2018;Zhu et al 2019;Zhang, Zhang, and Cheng 2020;Zhang and Zitnik 2020;Suresh et al 2021b). The most related to our work is ProGNN (Jin et al 2020) which explores the low-rank and sparsity of the graph structure, and SimP-GCN (Jin et al 2021b) which relies on a similarity preservation scheme for structure learning.…”

“…Baselines: We follow the previous works (Jin et al 2021b;He et al 2022) to use eleven baselines. We categorize these methods into three groups: 1) multi-hop-based approaches MixHop (Abu-El-Haija et al 2019) and H2GCN (Zhu et al 2020), which mix the multi-hop neighbors for aggregation; 2) ranking-based approaches NLGNN (Liu, Wang, and Ji 2021), GEOM-GCN (Pei et al 2020), Node2Seq (Yuan and Ji 2021) and GPNN (Yang et al 2022) that aim to search on the network structure and then perform selective aggregation; 3) structure learning approaches ProGNN (Jin et al 2020), UGCN (Jin et al 2021a), BM-GCN (He et al 2022) and GREET (Liu et al 2023) that automatically learn graph structures for aggregations. Specifically, ProGNN preserves the low-rank and sparsity characteristics of the graph structure for robust GCN.…”

Federated Learning on Non-IID Graphs via Structural Knowledge Sharing

“…In the meantime, to capture global structural knowledge, we introduce a random walk-based structure embedding (RWSE) which is computed based on the random walk diffusion process (Tong, Faloutsos, and Pan 2006;Dwivedi et al 2021;Liu et al 2023b). Concretely, RWSE is denoted as…”