💻 computer science

XGNN-ID: Explainable Graph Neural Network Framework for Imbalanced Datasets

This paper introduces XGNN-ID, an explainable Graph Neural Network framework that integrates a novel Selective Validation-Weighted Ensemble with NodeImport Gating (SVW-NG) to effectively address class imbalance and enhance model interpretability across various graph datasets.

Original authors: Bhargavi B, Jaya Lakshmi Tangirala, Komal Ranjan Sahoo, Sai Sushanth G, Chellapuram Goutham Reddy, Akash Gundamaraju

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bhargavi B, Jaya Lakshmi Tangirala, Komal Ranjan Sahoo, Sai Sushanth G, Chellapuram Goutham Reddy, Akash Gundamaraju

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital world, data rarely comes in neat, balanced rows. Instead, it often arrives as a tangled web of connections, where one piece of information is linked to many others, much like a social network where a few famous people have thousands of friends while most have only a handful. To make sense of these complex webs, scientists use a type of artificial intelligence called a Graph Neural Network. These systems are designed to learn by looking at the relationships between things, allowing them to predict what a person might buy, identify a disease from a medical scan, or categorize a research paper by its topic. However, these powerful tools have a significant blind spot. They often struggle when the data they are fed is lopsided, with some categories appearing far more often than others. In such cases, the computer tends to ignore the rare, important details in favor of the common ones, leading to predictions that are accurate overall but unfair or useless for the minority groups. Furthermore, when these systems do make a decision, they often act like a black box, offering no explanation for why they chose one answer over another, which makes it difficult to trust them in high-stakes situations like healthcare or finance.

A team of researchers has developed a new framework to solve both of these problems at once. They created a system called XGNN-ID, which stands for an Explainable Graph Neural Network Framework for Imbalanced Datasets. The goal was to build a method that not only helps the computer learn fairly from unbalanced data but also opens the black box to show exactly how it reached its conclusions. The researchers tested their approach on a variety of real-world networks, including citation graphs where scientific papers link to one another and product networks where items are frequently bought together. They found that by combining several different strategies to balance the data and then carefully selecting the best ones, they could significantly improve the accuracy of the predictions, especially for the rare and often overlooked categories. More importantly, they discovered that while traditional balancing methods often caused the computer's reasoning to become erratic or degrade in quality, the new framework maintained a stable and dependable level of explanation quality, avoiding the sharp drops in reliability seen with other techniques.

The researchers began by acknowledging that standard methods for fixing unbalanced data often fail when applied to these complex webs. Techniques that work well for simple lists of numbers can break the delicate structure of a graph, causing the computer to lose important context. To address this, the team built a pipeline that first analyzes the graph to understand how uneven the data is. They then applied a variety of balancing techniques, such as giving more weight to the rare examples or creating synthetic copies of them to fill the gaps. Instead of relying on just one of these methods, they introduced a smart selection process that acts like a committee. This committee trains multiple versions of the model using different balancing tricks and then weighs their final predictions based on how well each version performed on a test set. This approach, which they call a selective validation-weighted ensemble, allows the system to pick the most reliable strategies for each specific situation, rather than forcing a single method to work for every problem.

When they put this framework to the test, the results were striking. On datasets representing scientific papers, the new method consistently outperformed traditional approaches. For instance, on a network of computer science papers, the system improved its ability to correctly identify the less common research topics by a significant margin, boosting the accuracy for these rare groups by over four percent compared to the baseline. On even larger and more complex networks, such as those representing products on an online shopping site, the improvement was dramatic. In one case involving computer hardware, the system's ability to correctly identify the rare items jumped by more than five hundred percent compared to the unbalanced starting point. These gains were not limited to just one type of model; the framework worked effectively across several different architectures, proving that the solution was robust and adaptable to various ways of processing graph data.

Perhaps just as important as the improved accuracy was the stability of the explanations. The researchers used a tool designed to highlight the specific parts of the graph that influenced a decision, essentially asking the computer to point to the evidence it used. They found that while some traditional balancing methods made the computer's reasoning erratic or confusing—sometimes causing noticeable degradations in explanation quality—the new framework kept the explanations steady. The system avoided the sharp fidelity drops observed with other techniques, maintaining a balanced and dependable explanation quality even as it learned from the corrected data. This means that when the model predicts a rare event, it does so with a clear and trustworthy rationale, rather than a guess based on noise. The study suggests that by carefully managing how data is balanced, it is possible to create artificial intelligence that is not only more accurate for everyone but also more transparent and easier to understand, bridging the gap between raw predictive power and human trust.

Technical Summary: XGNN-ID Framework for Imbalanced Graph Data

Problem Statement

Graph Neural Networks (GNNs) have demonstrated robust predictive capabilities in graph-structured data domains such as citation networks, fraud detection, and recommendation systems. However, two critical challenges hinder their deployment in high-stakes environments:

Opacity: Most GNN models function as "black boxes," lacking transparency in their decision-making processes, which is a barrier in fields like healthcare and finance where interpretability is paramount.
Class Imbalance: Real-world graph datasets often suffer from severe class imbalance, where minority classes are under-represented. Traditional GNNs trained on such data tend to bias towards majority classes, leading to poor generalization, unfair predictions, and unstable behavior.

While class rebalancing techniques (e.g., Weighted Loss, Focal Loss, SMOTE) and explainability tools (e.g., GNNExplainer) exist independently, the intersection of these domains remains underexplored. Specifically, there is a lack of systematic analysis regarding how balancing strategies affect not only predictive accuracy but also the quality, stability, and coverage of GNN explanations.

Methodology: The XGNN-ID Framework

The authors propose XGNN-ID (Explainable Graph Neural Network Predictions for Imbalanced Datasets), a unified framework designed to evaluate the interplay between class balancing strategies and explainability techniques across multiple GNN architectures.

Core Components

The framework integrates three primary modules:

Class Balancing Module: This module modulates the learning process using various techniques, including:
- Weighted Cross-Entropy Loss: Assigns higher importance to underrepresented classes.
- Focal Loss: Prioritizes hard-to-classify samples by reducing the loss contribution from well-classified instances.
- Oversampling/Undersampling: Adjusts class distributions via data augmentation or reduction.
- NodeImport: A meta-set guided framework that assesses node importance to selectively filter error-prone majority samples and promote informative minority samples.
- SVW-NG (Selective Validation-Weighted Ensemble with NodeImport Gating): A novel ensemble technique proposed by the authors. It trains multiple base models using different balancing strategies and aggregates their predictions using validation-guided weights. The weight $\alpha_k$ for each model $m_k$ is determined by its validation Macro-F1 score ( $s_k$ ), ensuring that models performing well on minority classes receive stronger emphasis.
Model Training and Regularization Module: This component ensures stability during training through standard regularization techniques, specifically Dropout (to prevent overfitting by randomly turning off neurons) and Weight Decay (L2 regularization to encourage smoother convergence).
Explainability Module: Utilizes GNNExplainer as a post-hoc analysis tool to generate node-level and model-level explanations. It learns masks on graph structures (edges) and node features to identify the most informative components influencing predictions. The quality of these explanations is evaluated using Fidelity (how accurately the explanation replicates the original prediction) and Sparsity (the degree to which the explanation condenses the graph by removing redundant edges).

Experimental Setup

The framework was evaluated on five benchmark datasets:

Citation Networks: Cora, CiteSeer, and PubMed.
Product Co-purchase Networks: Amazon-Computers and Amazon-Photo.

Five foundational GNN architectures were tested: GCN, GraphSAGE, GAT, ChebNet, and GIN. The experiments compared the proposed SVW-NG ensemble against baseline training and individual balancing techniques (Weighted, Focal, Oversampling, Undersampling, GraphSMOTE, NodeImport).

Key Contributions

Unified Framework (XGNN-ID): A systematic pipeline that combines class imbalance correction with post-hoc explainability analysis, allowing for the simultaneous assessment of predictive performance and interpretability.
SVW-NG Ensemble: The introduction of a validation-weighted ensemble method that selectively combines balancing techniques. This approach leverages the strengths of individual methods (e.g., NodeImport for noise filtering) while mitigating their weaknesses through weighted aggregation based on validation performance.
Cross-Model Empirical Study: A comprehensive analysis of how different balancing mechanisms influence accuracy, stability, and training dynamics across diverse GNN architectures.
Explainability–Imbalance Interaction Analysis: An assessment of how balancing strategies impact explanation quality (fidelity and sparsity), demonstrating that balancing does not merely affect accuracy but also alters the interpretability profile of the resulting GNN.

Results

Extensive experiments on the benchmark datasets yielded the following findings:

Performance Improvement: The proposed SVW-NG strategy generally outperformed baseline models and individual balancing techniques, though it did not strictly dominate every architecture-dataset combination.
- On CiteSeer, SVW-NG achieved up to a 4.18% improvement in Balanced Accuracy and 4.38% in Macro-F1 for the GIN model compared to the baseline.
- On Amazon-Computers, the improvements were particularly stark for GCN, with a 566% increase in Balanced Accuracy and a 1045% increase in Macro-F1 over the baseline.
- SVW-NG consistently delivered competitive or superior performance compared to the strongest existing balancing methods (e.g., outperforming the best competing method by 5.13% in Macro-F1 for GCN on Amazon-Photo).
- Note: While SVW-NG showed significant gains in most cases, it did not outperform the baseline in all scenarios; for instance, on the Cora dataset, the GAT model achieved a slightly higher Balanced Accuracy with the baseline (0.827) compared to SVW-NG (0.823).
Explainability Stability:
- Fidelity: While traditional methods (especially oversampling) often caused fidelity degradation (ranging from -3% to -8%), XGNN-ID with SVW-NG demonstrated consistently stable fidelity across models and datasets. It improved fidelity by approximately 2% to 5% on citation and Amazon datasets for GCN and ChebNet, avoiding the sharp drops seen in other techniques.
- Sparsity: The framework maintained high sparsity scores (close to 1), indicating that the explanations remained clean and focused on influential relationships without excessive noise.
Ablation Study: The inclusion of SVW-NG led to consistent improvements in Balanced Accuracy and Macro-F1, particularly on sparse graphs like Cora and CiteSeer, confirming the efficacy of the ensemble approach.

Significance and Claims

The paper posits that XGNN-ID bridges a critical gap in graph machine learning by addressing the dual challenges of class imbalance and model opacity. The authors claim that:

Balancing Enhances Interpretability: The study demonstrates that balancing methods do not only improve predictive accuracy but can also stabilize and enhance the quality of explanations, making GNNs more trustworthy.
Robustness in Imbalanced Settings: The framework identifies models and strategies that are most robust when dealing with imbalanced data, providing a pathway for deploying GNNs in high-risk domains where fairness and reliability are essential.
Unified Evaluation: By straddling the boundaries of predictive modeling and rational assessment, XGNN-ID offers a foundation for developing transparent, fair, and trustworthy graph learning systems.

The authors conclude that while the framework shows promise, future work should extend these evaluations to larger, industrial-scale datasets and explore high-order connectivity representations in dynamic and heterogeneous networks.