Abstract
Multi-view Graph Neural Networks (GNNs) typically operate under the assumption that features consistent across multiple views are the most informative for downstream tasks. However, this "consistency" paradigm often captures "incidental factors"—spurious correlations that are consistent across views but irrelevant to the actual target label. We propose the Target-oriented Graph Neural Network (TGNN), which shifts the focus from simple cross-view consistency to target-oriented alignment. By disentangling determinative (causal) features from incidental noise through a dual-objective loss function, TGNN achieves superior generalization and robustness. Experimental results across nine datasets demonstrate that TGNN consistently outperforms state-of-the-art baselines.
The Consistency Trap in Multi-view Learning
In multi-view learning, the prevailing philosophy is to seek consensus among different data sources. While intuitive, this approach is vulnerable to the "consistency trap." For instance, in a dataset of horse images taken from various angles, "green grass" might appear consistently across all views. A standard model might erroneously identify grass as a defining feature of a horse. These incidental factors lead to poor generalization when the model encounters a horse in a different environment, such as a stable.
Existing GNNs for multi-view data often fail to distinguish between these causal (determinative) and noisy (incidental) semantics. Furthermore, many current models lack a rigorous theoretical foundation to explain their performance on unseen data.
The TGNN Framework
The proposed TGNN architecture addresses these challenges through three primary mechanisms:
1. Semantic Disentanglement
TGNN employs a dual-extractor strategy for each view. A shared determinative extractor ($CE$) focuses on capturing the core causal relationships necessary for the target task. Simultaneously, a view-specific incidental extractor ($SE^v$) is trained to identify and isolate biases and noise inherent to that specific view.
2. Class-level Dual Objective (CDO) Loss
The model is optimized using a "push-pull" loss strategy:
- The Pull (Alignment): A standard cross-entropy loss ($L_{ce}$) ensures that the determinative features are highly predictive of the ground-truth labels.
- The Push (Separation): A prediction separation loss ($L_{sep}$) forces the representations produced by the incidental extractor to be as distinct as possible from those of the determinative extractor. This is reinforced by a semantic separability penalty that maintains a physical distance between these representations in the latent space.
3. Theoretical Generalization Guarantees
TGNN is backed by a formal generalization analysis. The authors derive an upper bound on the loss discrepancy between training and test distributions using KL divergence. This provides a mathematical guarantee that minimizing empirical risk on training data translates to robust performance in unknown test domains.
Experimental Evaluation
The researchers evaluated TGNN against 16 state-of-the-art competitors (including HAN, DMGI, and ECMGD) across nine diverse datasets, spanning multi-relational, multi-attribute, and multi-modal categories.
Key Results
- Performance: TGNN achieved the highest scores in Accuracy and Macro-F1 across nearly all datasets. For example, on the ACM dataset, it reached a Macro-F1 of 94.09%, significantly outperforming the next best competitor.
- Adversarial Robustness: When tested against Mettack (an adversarial attack perturbing 25% of graph edges), TGNN exhibited significantly higher stability than other GNNs. By ignoring incidental factors, the model effectively filters out malicious noise.
- Ablation Studies: Removing either the KL divergence term or the separation loss resulted in a measurable drop in performance, validating the necessity of the model's theoretical components.
Conclusion
TGNN represents a significant shift in multi-view graph learning by moving beyond the limitations of the consistency paradigm. By explicitly disentangling determinative features from incidental noise, the model achieves higher accuracy and greater robustness. While future work may address hyperparameter sensitivity and extension to dynamic graphs, TGNN currently sets a new benchmark for target-oriented multi-view representation learning.
Reference: Huang, S., Fu, L., et al. (2025). Going Beyond Consistency: Target-oriented Multi-view Graph Neural Network. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence (IJCAI-25).
Summary updated