RA-GCN: Graph Convolutional Network for Disease Prediction Problems with Imbalanced Data (2103.00221v3)

Abstract: Disease prediction is a well-known classification problem in medical applications. GCNs provide a powerful tool for analyzing the patients' features relative to each other. This can be achieved by modeling the problem as a graph node classification task, where each node is a patient. Due to the nature of such medical datasets, class imbalance is a prevalent issue in the field of disease prediction, where the distribution of classes is skewed. When the class imbalance is present in the data, the existing graph-based classifiers tend to be biased towards the major class(es) and neglect the samples in the minor class(es). On the other hand, the correct diagnosis of the rare positive cases among all the patients is vital in a healthcare system. In conventional methods, such imbalance is tackled by assigning appropriate weights to classes in the loss function which is still dependent on the relative values of weights, sensitive to outliers, and in some cases biased towards the minor class(es). In this paper, we propose a Re-weighted Adversarial Graph Convolutional Network (RA-GCN) to prevent the graph-based classifier from emphasizing the samples of any particular class. This is accomplished by associating a graph-based neural network to each class, which is responsible for weighting the class samples and changing the importance of each sample for the classifier. Therefore, the classifier adjusts itself and determines the boundary between classes with more attention to the important samples. The parameters of the classifier and weighting networks are trained by an adversarial approach. We show experiments on synthetic and three publicly available medical datasets. RA-GCN demonstrates the superiority compared to recent methods in identifying the patient's status on all three datasets. The detailed analysis is provided as quantitative and qualitative experiments on synthetic datasets.

• The paper presents RA-GCN, a novel adversarial graph convolutional network that dynamically re-weights samples to counter class imbalance in disease prediction.
• It integrates separate weighting networks with a GCN classifier to optimize weighted cross-entropy loss and enhance accuracy on minority classes.
• Extensive validation on medical datasets shows RA-GCN outperforms traditional approaches with improved accuracy, macro F1-score, and ROC AUC metrics.

RA-GCN: Graph Convolutional Network for Disease Prediction Problems with Imbalanced Data

The paper presents a novel method, the Re-weighted Adversarial Graph Convolutional Network (RA-GCN), designed to address the challenge of class imbalance in disease prediction using Graph Convolutional Networks (GCNs). The method particularly focuses on enhancing the predictive performance of GCNs in medical datasets characterized by class imbalance, where the number of samples in different classes is significantly disproportionate. Such imbalance often leads to models that favor the majority class, thereby reducing the accuracy in identifying minority classes, which are frequently the cases of interest in medical applications.

Methodology

RA-GCN innovatively combines GCNs with adversarial training, employing separate graph-based neural networks for each class to dynamically assign weights to samples during the training process. This adaptive weighting mechanism prevents the classifier from overemphasizing samples from any particular class, which is a common issue when class imbalance is present. The proposed model incorporates two main components: a GCN-based classifier responsible for node classification, and a series of weighting networks, each focusing on learning the appropriate sample weights for one class.

The adversarial nature of the training process involves the classifier minimizing the weighted cross-entropy loss, while the weighting networks are designed to maximize the importance of the samples that are more challenging for the classifier. This adversarial setup enables the system to learn a balanced classification boundary, addressing biases introduced by class imbalance.

Experimental Validation and Results

The authors validate RA-GCN on both synthetic data and real-world medical datasets, including the Pima Indian Diabetes dataset, Parkinson's Progression Markers Initiative dataset, and the Haberman's survival dataset. Across these datasets, RA-GCN consistently demonstrates superior performance compared to both weighted and unweighted versions of traditional GCNs and Multi-Layer Perceptrons (MLPs), as well as a state-of-the-art method for handling class imbalance in GCNs, referred to as DR-GCN.

The results are measured in terms of accuracy, macro F1-score, and ROC AUC, which collectively offer a comprehensive evaluation of the classifier's performance across both majority and minority classes. RA-GCN achieves higher metrics by effectively discriminating between poorly represented classes while maintaining high overall accuracy. This success is attributed to its capability to balance the classes during training dynamically.

Implications and Future Directions

The development of RA-GCN has significant implications for machine learning applications in healthcare, particularly in tasks involving critical but low-prevalence conditions where it is essential to minimize false negatives. Beyond healthcare, RA-GCN's adversarial sample weighting strategy could be adapted to other domains suffering from class imbalance, such as fraud detection and anomaly detection.

Future research could explore the extension of RA-GCN to multi-label classification scenarios, seek methods to integrate this approach with other GCN architectures and implement further optimization for training on very large graphs. Furthermore, the inductive learning capabilities, which allow for predictions on unseen nodes, could be a valuable avenue, given the transductive nature of the current model setting. These directions promise to enhance the utility and robustness of GCNs in various imbalanced-class scenarios.