Abstract
Heterogeneous Information Networks (HINs) provide a rich representation of data by capturing multiple types of nodes and relations. However, the high cost of labeling HINs limits the applicability of supervised learning. While Domain Adaptation (DA) offers a solution by transferring labels from a source to a target domain, HINs introduce unique challenges: semantic complexity and the "similarity-distinguishability trap," where aligning domains can inadvertently suppress discriminative features. This paper introduces MuSDAC (Multi-space Domain Adaptive Classification), a framework that leverages multiple embedding spaces to maintain semantic richness while achieving domain alignment. Through a multi-channel GCN architecture and a two-level selection strategy involving heuristic sampling and moving-averaged voting, MuSDAC achieves superior performance on cross-domain classification tasks compared to existing state-of-the-art methods.
Introduction
Heterogeneous Information Networks (HINs) are powerful tools for modeling complex real-world systems, such as academic citation networks where nodes represent authors, papers, and venues connected by various semantic relations (meta-paths). Despite their utility, the scarcity of labeled data in many target domains necessitates effective Domain Adaptation (DA) techniques.
Standard DA approaches typically aim for domain invariance—making source and target distributions indistinguishable. However, in the context of HINs, this often leads to two major issues. First, squashing the multi-layered semantics of an HIN into a single global space often results in significant information loss. Second, features that are most invariant across domains are frequently the least useful for classification, a phenomenon termed the "similarity-distinguishability trap." MuSDAC is designed to overcome these hurdles by performing alignment across multiple semantic spaces independently.
Methodology
The MuSDAC framework consists of three primary components: Multi-channel Shared Weight GCNs, Multi-space Alignment, and a Two-level Selection Strategy.
Multi-channel Shared Weight GCNs
To capture the diverse semantics of HINs, the network is decomposed into multiple channels based on specific meta-paths. Each channel utilizes a Graph Convolutional Network (GCN) with shared weights between the source and target domains. This ensures that the initial embeddings for both domains reside in a comparable feature space before further alignment.
Multi-space Alignment
Instead of a single global alignment, MuSDAC performs pairwise alignment across multiple embedding spaces. By using Maximum Mean Discrepancy (MMD) in these distinct spaces, the model can preserve specific semantic relationships that might be lost in a unified space. This approach allows the model to sidestep "trap-vectors"—features that increase domain similarity at the expense of classification performance.
Two-level Selection Strategy
Managing multiple channels requires an efficient selection and fusion mechanism:
- Heuristic Combination Sampling: To avoid the exponential complexity of testing all channel combinations, a polynomial-time heuristic identifies combinations that maximize both transferability and distinguishability.
- Moving-averaged Weighted Voting: The final classification is a fusion of outputs from different spaces. A moving average is applied to the voting weights to prevent "Weight Dominance," ensuring stable convergence across all semantic channels.
Experiments and Results
MuSDAC was evaluated on three academic citation datasets: ACM, AMiner, and DBLP. The experiments focused on challenging cross-domain scenarios where structural differences between networks were significant.
The results indicate that MuSDAC consistently outperforms baseline models such as GCN, HAN (Heterogeneous Attention Network), and DANE (Domain Adaptive Network Embedding). MuSDAC achieved an average accuracy of 0.781, significantly higher than DANE (0.688) and HAN (0.715). Ablation studies confirmed the necessity of the two-level selection strategy; replacing it with simple averaging or random selection led to a noticeable drop in accuracy (e.g., from 0.726 to 0.702 in the ACM A→B task).
Key Insights
A central finding of this work is that the upper bound of prediction error on a target domain is lower when minimizing the weighted sum of losses across multiple embedding spaces rather than a single one. Furthermore, the moving-average voting scheme proved essential for stabilizing gradients in a multi-channel architecture, allowing diverse semantic signals to contribute effectively to the final prediction.
Conclusion
MuSDAC provides a robust solution for domain adaptive classification on HINs by embracing semantic diversity through multi-space alignment. By avoiding the "similarity-distinguishability trap," the framework maintains high classification accuracy even in the presence of significant domain shifts. Future work will explore automated meta-path discovery and the integration of adversarial training to further enhance the framework's flexibility and performance.
Summary updated