FAKER: Generating Frequency-based Artificial Attributes via Random Walks for Non-attribute Graph Representation Learning

Mincheol Shin, Taeyoung Choe, YEJONG RYU, Yunyong Ko, Mucheol Kim

Keywords: graph neural network, attribute-missing, random walk, artificial node features, non-attributed graph

Abstract: A key challenge for Graph Neural Networks (GNNs) is their reliance on initial node features, while many real-world graphs lack such attributes due to privacy constraints or limitations in data collection. Existing adjacency-only approaches attempt to learn representations directly from topology. However, they often inherit the sampling biases of random walks, leading to skewed embeddings. To address these limitations, we propose FAKER, a diagnosis‑driven, model‑agnostic framework that synthesizes artificial node attributes from topology alone. FAKER first analyzes group-level visit signals from random walks with Power Spectral Density (PSD) to quantify low-frequency persistence bias and high-frequency switching bias. The resulting quantified scores then drive an adaptive sampler that produces a balanced corpus without distributional assumptions by reweighting transitions and allocating additional walks. A lightweight co‑occurrence encoder trained on this corpus yields dynamic features, which are merged with a compact structural summary and standardized to form plug‑and‑play attributes for any GNN. Across four benchmarks, FAKER establishes state-of-the-art results among adjacency-only baselines for node classification and link prediction. It also matches or outperforms feature-using methods on three datasets. Ablation and robustness studies show that the improvements result from the frequency-domain diagnosis and adaptive allocation, rather than from the number of walks or sensitive hyperparameter tuning.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 24244