TL;DR
This paper introduces CSGNN, a novel framework for robust node classification on graphs with noisy labels and class imbalance. CSGNN dynamically selects reliable nodes by considering class-specific confidences, outperforming existing methods by effectively mitigating label noise and class imbalance simultaneously.
What Problem Does This Paper Address?
Graph Neural Networks (GNNs) are powerful tools for learning representations from graph-structured data, finding applications in diverse fields like social networks and knowledge graphs. However, GNNs often struggle with two critical issues: noisy labels (errors in the ground truth labels used for training) and class imbalance (where some classes have significantly fewer samples than others).
Noisy labels can lead GNNs to learn incorrect patterns, especially in graphs where label noise can propagate through edges, compounding errors. Class imbalance exacerbates this problem, as standard training procedures tend to favor majority classes, leading to poor performance on minority classes. Existing methods for handling noisy labels often fail to address class imbalance, or vice-versa, and many are designed for independent and identically distributed (i.i.d.) data, making them ill-suited for the interconnected nature of graphs. This paper aims to bridge this gap by proposing a unified approach that tackles both noisy labels and class imbalance in graph node classification.
Methods
CSGNN tackles the problem through a multi-stage approach:
- Graph Augmentation: To address the sparsity of real-world graphs and improve neighbor aggregation, CSGNN first augments the graph. It trains an "edge predictor" GNN to learn a latent representation of nodes and then uses cosine similarity between these representations to predict potential edges between labeled and unlabeled nodes. This densifies the graph, allowing for better information propagation.
- Label Aggregation: To obtain more reliable confidence measures than single-node predictions, CSGNN aggregates information from neighboring nodes. It uses the augmented graph and a GNN encoder to project nodes into a latent space. Then, for each node, it calculates the K' nearest neighbors in this latent space and aggregates their representations to derive a class-wise confidence distribution. This process is further refined by reweighting the aggregation prediction by the class prior to mitigate the impact of class imbalance.
- Dynamic Class-wise Selection: This is the core innovation of CSGNN. Instead of using a global threshold for selecting clean nodes (which can lead to over- or under-sampling of certain classes), CSGNN employs a dynamic, class-wise selection mechanism. It uses K-Means clustering on the aggregated confidence scores within each class to identify a clean subset of nodes. This ensures that nodes from both majority and minority classes are selected based on their actual reliability, effectively addressing class imbalance.
- Robust Training: The selected clean nodes are then used to train a GNN classifier using a cross-entropy loss with label smoothing. Label smoothing acts as a regularizer, preventing the model from overfitting to the potentially noisy labels within the selected clean set and improving generalization.
Datasets
The paper evaluates CSGNN on three widely-used benchmarks:
- Cora: A citation network dataset.
- Citeseer: Another citation network dataset.
- BlogCatalog: A social network dataset.
These datasets are split into training, validation, and testing sets with standard ratios (5%, 15%, 80%). Label noise (uniform and pair noise) is artificially introduced into the training and validation sets at various ratios (10%, 20%, 30%, 40%).
Key Insights and Ideas
- Class Imbalance Amplifies Label Noise Issues: The paper highlights that class imbalance isn't just a separate problem; it actively worsens the impact of label noise. A global threshold for noise filtering will disproportionately affect minority classes, leading to biased learning.
- Neighbor Aggregation is Key for Reliable Confidence: Relying on a single node's confidence is insufficient. Aggregating information from neighbors in a learned latent space provides a much more robust measure of a node's true label.
- Class-wise Clustering is Crucial for Fairness: The dynamic, class-specific clustering approach is the linchpin for handling both noise and imbalance. It allows the model to adaptively identify clean samples from each class, preventing the bias inherent in global methods.
- Graph Augmentation Enhances Information Flow: Densely connecting labeled and unlabeled nodes helps GNNs leverage more information and improve their ability to learn robust representations, especially when dealing with limited labeled data.
Key Results
CSGNN consistently outperforms state-of-the-art methods across all tested benchmarks and noise settings.
- On the Cora dataset with 40% uniform noise, CSGNN achieves an accuracy of 76.6%, significantly higher than the next best method (NRGNN at 70.5%).
- Similar improvements are observed for Citeseer and BlogCatalog, particularly under higher noise rates.
- The ablation studies confirm the effectiveness of each component: graph augmentation, label aggregation, and class-wise selection all contribute significantly to CSGNN's performance. Replacing class-wise selection with a fixed threshold leads to a substantial performance drop, especially on imbalanced datasets.
Limitations and Future Work
- Correction Module: The authors suggest that incorporating a dedicated "correction module" could further improve the utilization of the entire dataset, potentially by refining the labels of even the selected "clean" nodes.
- Stability at High Noise Rates: While CSGNN performs well, the paper notes that its stability can decrease at very high noise rates, indicating that extreme noise remains a challenge.
- Extension to i.i.d. Data and Prior Knowledge: Future work could explore extending CSGNN's principles to i.i.d. data and integrating external prior knowledge, perhaps from large language models, to further mitigate bias.
Summary updated