Adversarial Graph Disentanglement

TL;DR

ADGCN addresses graph disentanglement by combining micro-disentanglement through component-specific aggregation with macro-disentanglement via a conditional adversarial regularizer. It also introduces a diversity-preserving, locally refined graph construction to reveal latent structure progressively. Empirical results on eight real-world graphs show competitive or superior performance in node classification and clustering, with clear evidence of disentanglement and robustness gains. The combination of micro- and macro-disentanglement, together with progressive graph refinement, provides improved interpretability and resilience for graph representations in real-world tasks.

Abstract

A real-world graph has a complex topological structure, which is often formed by the interaction of different latent factors. However, most existing methods lack consideration of the intrinsic differences in relations between nodes caused by factor entanglement. In this paper, we propose an \underline{\textbf{A}}dversarial \underline{\textbf{D}}isentangled \underline{\textbf{G}}raph \underline{\textbf{C}}onvolutional \underline{\textbf{N}}etwork (ADGCN) for disentangled graph representation learning. To begin with, we point out two aspects of graph disentanglement that need to be considered, i.e., micro-disentanglement and macro-disentanglement. For them, a component-specific aggregation approach is proposed to achieve micro-disentanglement by inferring latent components that cause the links between nodes. On the basis of micro-disentanglement, we further propose a macro-disentanglement adversarial regularizer to improve the separability among component distributions, thus restricting the interdependence among components. Additionally, to reveal the topological graph structure, a diversity-preserving node sampling approach is proposed, by which the graph structure can be progressively refined in a way of local structure awareness. The experimental results on various real-world graph data verify that our ADGCN obtains more favorable performance over currently available alternatives. The source codes of ADGCN are available at \textit{\url{https://github.com/SsGood/ADGCN}}.

Figures (14)

• Figure 1: A simple example in a citation network. A paper $u$ for visual question answering (VQA) may simultaneously have citation relationships with papers in the fields of ML, NLP, and CV because of the related researches it involves. Through differentiated aggregation, the learned separable representation by graph disentanglement can reveal more about the latent semantics among different neighborhoods.
• Figure 2: Architecture overview of a single-layer ADGCN. $A^{(t)}$ represent the refined adjacency at epoch $t$. $\mathbf{X}$ and $\mathbf{W}_{k}$ denote the node raw feature and the projection matrix to $k$-th component space respectively. $\{\mathbf{h}^{k}_{u}\}^{K}_{k=1}$ represent the components of node $u$ and $\mathbf{h}^{(t)}_{u}$ as the output denotes the disentangled representations at epoch $t$, where $\mathbf{h}^{(t)}_{u}$ is used to update $A^{(t)}$ to $A^{(t+1)}$ through sample-based graph refinement.
• Figure 3: Illustration of component-specific aggregation based on dynamic assignment. The proposed dynamic assignment iteratively allocates the neighbor nodes to different component spaces according to the hubness score $p^{k}_{u \to v}$. Meanwhile, according to the authority score $q^{k}_{v \to u}$ of node $v$ in each component space, $\{\mathbf{c}^{k}_{i}\}_{i \in \tilde{\mathbf{N}}(u)}$ are aggregated into $\tilde{\mathbf{h}}^{k}_{u}$. Here, only three component spaces are considered.
• Figure 4: The iterative adversarial training tends to progressively improve the separability among component distributions.
• Figure 5: Illustration of the diversity-preserving graph refinement. Since the initially given adjacency matrix $\mathbf{A}$ is extremely sparse, we have marked the existing edges $\mathbf{A}_{ij}$ using a circle for clearer presentation.