PALA: Class-imbalanced Graph Domain Adaptation via Prototype-anchored Learning and Alignment
Graph Domain Adaptation (GDA) is a critical task for node classification when labeled data exists for a "source" graph (e.g., an older citation network) but labels are missing for a "target" graph (e.g., a modern network). A major real-world hurdle is that graphs are almost always class-imbalanced: a few "head" classes have many nodes, while many "tail" classes have very few. Standard GDA models often overfit to these head classes, leading to biased knowledge transfer and poor performance on rare but important tail classes.
The PALA Framework
PALA (Prototype-anchored Learning and Alignment) addresses these challenges through a three-pronged strategy designed to tackle imbalance and domain shift simultaneously.
1. Graph Consistency Encoder
Instead of relying solely on immediate neighbors, PALA utilizes Positive Pointwise Mutual Information (PPMI). This allows the encoder to capture high-order topological relationships (global consistency), creating more generalized node representations that are more invariant across different domains of the same type.
2. Prototype-anchored Recalibration & Learning
Rather than comparing individual nodes, which can be noisy in imbalanced settings, the model learns a single categorical prototype (a representative vector) for each class.
- Adversarial Proto-instances: To help the model distinguish between similar classes when data is scarce, PALA synthesizes "hard" negative examples by perturbing tail-class features with head-class prototypes.
- Recalibration: It uses an Exponential Moving Average (EMA) to calculate a calibration factor, ensuring that tail-class prototypes remain influential and aren't "drowned out" by majority classes during training.
3. Contrastive Prototype Adaptation
To bridge the gap to the target graph, PALA generates pseudo-labels for target nodes. It then uses a weighted contrastive loss to align these target features with the refined source prototypes, focusing more on high-confidence target samples to reduce noise.
Key Insights and Results
- Prototypes as Stable Anchors: Using class-level representations is significantly more robust than node-to-node matching when dealing with imbalanced data. It provides a stable "target" for alignment that isn't as sensitive to the lack of individual samples in tail classes.
- Superiority over SOTA: Evaluated on real-world citation networks (ACMv9, Citationv1, DBLPv7), PALA consistently outperformed both standard GNNs and specialized domain adaptation methods. Even at an Imbalance Factor of 50 (where the largest class is 50x larger than the smallest), PALA maintained high Macro-F1 scores.
- Fixing the Source is Critical: Ablation studies reveal that addressing imbalance in the source domain (via logit compensation and prototype learning) is more critical than the domain alignment itself. If the source model is biased, alignment simply transfers that bias to the target.
Limitations and Future Directions
While effective, PALA faces challenges in scaling global PPMI calculations to massive, billion-node graphs. Additionally, the sensitivity analysis showed that the parameter $\gamma$ (related to adversarial samples) requires careful tuning to avoid amplifying imbalance. Future work may involve integrating Large Language Models (LLMs) into the prototype-anchored framework to leverage semantic text data for even better cross-domain generalization.
Summary updated