GNNmim: Graph Neural Networks with Missing Indicators

• GNNmim is a model that integrates explicit missing indicator masks with zero-imputed node features to enhance GNN-based node classification under incomplete data.
• It employs standard GNN architectures by concatenating imputed features with binary masks, enabling effective handling of varied missingness mechanisms.
• Empirical evaluations show GNNmim achieves top F1 accuracy and minimal performance drop under MNAR and distribution shifts compared to baseline methods.

Graph Neural Networks with Missing Indicator Method (GNNmim) are a family of models designed for node classification on graphs subject to incomplete node feature information. The approach is predicated on integrating explicit binary masks of missing data with simple imputation, and then applying standard GNN architectures. GNNmim provides a robust, theoretical, and empirically validated baseline for dealing with missing node features, applicable to dense, real-world attributes and challenging missingness mechanisms beyond Missing Completely At Random (MCAR) (Ferrini et al., 8 Jan 2026).

1. Problem Context and Motivation

Node classification in attributed graphs is foundational across domains such as healthcare, sensor networks, and power systems. Unlike “academic” graph classification scenarios with dense feature coverage, real-world data commonly exhibit missing node features, with causes ranging from sensor dropout to privacy filtering. Conventional GNN benchmarks (e.g., CORA, CITESEER) are dominated by sparse bag-of-words attributes—creating an artificial robustness under MCAR. To move beyond this, research formalizes the incomplete data problem as learning a mapping

$P_\theta(Y|X_\mathrm{obs}, M)$

where $X$ is the (possibly incomplete) node-feature matrix and $M$ is the binary mask ($M_{ij}=1$ iff $X_{ij}$ is missing), for a set of nodes $V$ and edges $E$.

2. Missingness Mechanisms and Information-Theoretic Analysis

The missingness process is characterized by the probability model $P_a(M\,|\,X,Y)$. GNNmim rigorously distinguishes three key settings:

• Feature–MAR (Missing At Random with respect to $X$): $$P_a(M\,|\,X_\mathrm{obs}, X_\mathrm{miss}) = P_a(M\,|\,X_\mathrm{obs})$$
• Label–MAR: $P_a(M\,|\,X,Y) = P_a(M\,|\,X)$
• MCAR (Missing Completely At Random): $P_a(M\,|\,X,Y)$ is constant

Under MAR/MCAR, conditional distributions of $Y$ given $X_\mathrm{obs}, M$ are theoretically equivalent to those using $Y$ given mean- or zero-imputed $X_\mathrm{obs}$. The information loss for MCAR, especially under sparse features, is tightly bounded:

$-n d p\, h_2(s(X)) \le \Delta I \le 0$

where $h_2(u)$ is the binary entropy function and $s(X)$ is feature sparsity. For $s(X)\approx 1$, as in text graphs, information loss is negligible up to extremely high missingness, indicating that standard benchmarks do not probe the robustness of GNNs to missing data (Ferrini et al., 8 Jan 2026).

3. GNNmim Methodology

The GNNmim model implements the Missing Indicator Method (MIM):

• Imputation: All missing entries are zero-padded: $\tilde{X}_{ij} = X_{ij}$ if observed, $0$ if missing.
• Mask Concatenation: For each node $i$, the input becomes $[\tilde{x}_i \| m_i]$ where $m_i$ is the binary mask and “$\|$” denotes vector concatenation.
• GNN Stack: The augmented inputs are processed by a standard $L$-layer GNN (e.g., GCN, GraphSAGE, GIN) using typical message-passing operations. Mask bits provide the network architectural knowledge of missingness locations, allowing the model to “learn to ignore” zero placeholders.

The approach eschews learned or generative imputation. The training objective remains the standard cross-entropy for node classification:

$$L(\theta) = -\sum_{i \in V_L} \sum_{c=1}^C \mathbb{I}\{y_i=c\} \log \hat{y}_{i,c}$$

where $V_L$ is the set of labeled nodes, $C$ the class count, and $\hat{y}_{i}$ the softmax output.

4. Theoretical Properties

Under MCAR or MAR missingness, GNNmim is theoretically sufficient: including the mask yields no additional information if zero-imputation is unbiased. Under MNAR (Missing Not At Random), ignorability fails, and architectures that consume the mask $M$ as explicit input can, in principle, exploit missingness patterns to improve robustness or accuracy—though no formal theorem for universal sufficiency exists in the MNAR regime. Empirical robustness of GNNmim is observed across multiple MNAR scenarios, suggesting practical utility beyond the classical MCAR/MAR settings (Ferrini et al., 8 Jan 2026).

5. Empirical Evaluation and Robustness

Benchmarked across four dense-feature datasets (SYNTHETIC, AIR, ELECTRIC, TADPOLE) and five missingness mechanisms—including MCAR, MNAR (feature-dependent/class-dependent), and distribution-shifted train-test splits—the following trends are established:

• On BoW/TF-IDF graphs, all methods fail only at near-total missingness; such datasets provide little discriminatory power.
• For dense, semantically meaningful features, GNNmim consistently ranks within the top two approaches for node-level F1 (micro) accuracy across missingness regimes, with average F1-AUC values close to 0.80 (vs. 0.75–0.78 for alternatives).
• Under distribution shift (MNAR $\rightarrow$ MCAR in test), GNNmim shows the smallest performance drop among baselines.

Ablation confirms that supplementing competitor methods with the missingness mask generally provides 3–5 percentage points F1 gain, supporting the centrality of explicit missingness encoding. GNNmim scales to larger $n, d$ without loss of robustness.

6. Practical Usage and Recommendations

• Generalization: GNNmim is robust with no reliance on specific missingness assumptions or generative imputations, making it an assumption-free default for real-world settings with dense node features.
• Implementation Simplicity: The concatenation of mask and zero-imputed features incurs minimal computational cost and can be directly incorporated into any GNN architecture.
• Evaluation Practice: Researchers are strongly advised to benchmark exclusively on dense, semantically meaningful feature graphs and to subject models to diverse missingness and distribution shift scenarios for fair comparison.
• Baseline Recommendation: GNNmim should be the initial baseline prior to adopting more involved imputation or modeling techniques. For high-dimensional but sparse graphs, missingness robustness benchmarks are generally not meaningful.

7. Broader Implications and Limitations

GNNmim’s core principle—explicit mask concatenation with simple imputation—applies not just to node classification but, plausibly, to link prediction or other graph inference tasks under incomplete data regimes. However, when extreme MNAR or complex structural missingness artifacts are present, or when joint learning of explicit imputation or generative models is possible, GNNmim may be outperformed by architectures explicitly structured for such cases. Nevertheless, in a broad swath of realistic settings, especially in industrial or biomedical graph learning, GNNmim offers a robust, theoretically sound, and empirically validated approach (Ferrini et al., 8 Jan 2026).