Rethinking GNNs and Missing Features: Challenges, Evaluation and a Robust Solution

TL;DR

This paper tackles the practical challenge of missing node features in Graph Neural Networks by arguing that standard benchmarks with sparse features and MCAR missingness fail to stress-test robustness. It provides a theoretical bound showing information loss from missing features is limited under high sparsity and introduces dense, semantically meaningful datasets along with realistic missingness protocols, including LD-MCAR and MNAR variants with train–test shifts. The authors propose GNNmim, a simple Missing Indicator Method baseline that concatenates a missingness mask to the input features and uses a standard GNN, achieving competitive robustness across diverse missingness regimes without MAR assumptions. Overall, the work advocates for principled, realistic evaluation frameworks and demonstrates that lightweight approaches like GNNmim can outperform more complex architectures when evaluation artifacts are removed.

Abstract

Handling missing node features is a key challenge for deploying Graph Neural Networks (GNNs) in real-world domains such as healthcare and sensor networks. Existing studies mostly address relatively benign scenarios, namely benchmark datasets with (a) high-dimensional but sparse node features and (b) incomplete data generated under Missing Completely At Random (MCAR) mechanisms. For (a), we theoretically prove that high sparsity substantially limits the information loss caused by missingness, making all models appear robust and preventing a meaningful comparison of their performance. To overcome this limitation, we introduce one synthetic and three real-world datasets with dense, semantically meaningful features. For (b), we move beyond MCAR and design evaluation protocols with more realistic missingness mechanisms. Moreover, we provide a theoretical background to state explicit assumptions on the missingness process and analyze their implications for different methods. Building on this analysis, we propose GNNmim, a simple yet effective baseline for node classification with incomplete feature data. Experiments show that GNNmim is competitive with respect to specialized architectures across diverse datasets and missingness regimes.

Figures (8)

• Figure 1: Mean F1-score across 5 runs as a function of the missingness probability $\mu$ on the proposed datasets and established benchmarks. Each panel reports the performance of all models on a specific dataset under the S-MCAR setting. The complete tables for all missingness mechanisms are provided in Appendix \ref{['app:all_results']}.
• Figure 2: Column-normalized heatmaps showing the AUC (area under the F1 vs. missingness rate $\mu$ curve) for each model, dataset, and missingness mechanism. Higher values (lighter colors) indicate better overall robustness across increasing levels of missingness.
• Figure 3: F1 scores (mean ± std over 5 runs) under distribution shifts in missingness between training and test data. All models are trained with FD-MNAR missingness at 50%. Each panel corresponds to a dataset; each row to a model. Colored dots represent test-time F1 under U-MCAR with varying missingness rates: yellow = 0%, blue = 25%, green = 50%. Vertical red lines indicate the F1 achieved in the i.i.d. setting (FD-MNAR 50% at both train and test).
• Figure 4: F1 score as a function of feature missingness ($\mu$) for both classic benchmarks (top three rows) and our proposed datasets (bottom four rows), under the mechanisms described in Section \ref{['sec:missingness']}. Classic benchmarks show almost no degradation until extremely high $\mu$, while the proposed datasets reveal model weaknesses at more realistic missingness levels. Tables for numeric results are in App. \ref{['app:full_tables']}
• Figure 5: F1 score as a function of feature missingness ($\mu$) for additional synthetic datasets generated with the same procedure as Synthetic, but with either an increased number of nodes or features. For Synthetic4, the FairAC model is not reported since training exceeded the 12-hour time limit, while GOODIE is excluded due to out-of-memory errors.

Theorems & Definitions (8)

• Definition 1
• Theorem 1
• Definition 2: Feature Sparsity
• Theorem 2
• Theorem 2
• proof
• Theorem 2
• proof