TCHG: Tri-Trust Conditioned Heterogeneous Graph Learning for Reliable Dynamic Trust Prediction

Abstract: Trust prediction infers latent user-user trust relations and provides important support for social recommendation, fake-review and manipulation detection, and risk identification. Graph neural networks have become a prominent approach to trust prediction because of their ability to learn network structures and complex trust dependencies. However, existing methods often rely on a unified representation of trust signals and do not disentangle heterogeneous trust evidence into separate evidence channels, failing to exploit the distinct roles that different evidence channels should play during trust modeling. To address this gap, this paper argues that trust evidence should not be treated as an undifferentiated input, but should be decomposed and used as functional control factors over graph propagation. We propose TCHG, a tri-trust conditioned heterogeneous graph learning framework that decomposes trust evidence into three channels and assigns them distinct functional roles in propagation: entity reliability governs message admission, interaction-behavior reliability modulates propagation strength, and contextual trust adjusts the propagation mode through context-conditioned operator selection. Since the three evidence channels evolve at different temporal scales, TCHG maintains independent temporal states with non-uniform decay rates to prevent rapidly changing contextual signals from overwriting slowly accumulated entity reliability. It further predicts trust probability and calibrates the output probability, improving predictive confidence under sparse or conflicting evidence. Extensive experiments on multiple public trust datasets show that TCHG achieves effective and reliable trust prediction compared with representative trust prediction and heterogeneous graph baselines.

• The paper introduces a tri-trust evidence model that decomposes trust signals into entity, interaction-behavior, and contextual channels for refined graph propagation.
• It implements channel-specific propagation controls and component-decoupled temporal memory to enhance dynamic trust prediction and calibration.
• Empirical evaluations show superior MRR and robust uncertainty calibration across datasets, demonstrating improved resilience against poisoning attacks.

Tri-Trust Conditioned Heterogeneous Graph Learning for Reliable Dynamic Trust Prediction

Trust Prediction Problem and Existing Limitations

The paper "TCHG: Tri-Trust Conditioned Heterogeneous Graph Learning for Reliable Dynamic Trust Prediction" (2606.16611) addresses the challenge of modeling user–user trust relations in dynamic, heterogeneous web networks, a task relevant for recommendation, manipulation detection, and risk analyses. Traditional GNN-based trust models treat heterogeneous trust evidence as flattened features or generic attention weights, failing to exploit distinct roles across evidence channels (entity reliability, interaction-behavior reliability, contextual trust) in the propagation process. This conflation leads to ineffective message control and limits the reliability of learned trust predictions, especially under sparse or conflicting evidence.

TCHG Framework: Architecture and Propagation Mechanisms

TCHG introduces a structured evidence-controlled heterogeneous graph learning paradigm that decomposes trust signals into three functionally differentiated channels and assigns them distinct control roles during graph propagation: entity reliability governs message admission, interaction-behavior reliability modulates propagation strength, and contextual trust selects propagation operators conditioned on context.

Figure 1: A typical Web trust network where trust prediction relies on entity reliability, behavioral normality, and scenario-specific context evidence.

The workflow of TCHG consists of four modules:

1. Tri-Trust Evidence Construction and Encoding: Structured statistics are extracted from historical trust links, rating records, and local context, then encoded into latent embeddings and channel-specific reliability scores.
2. Tri-Trust Conditioned Heterogeneous Propagation: Entity reliability activates message-admission gates for each incoming edge; interaction-behavior reliability modulates message intensity; contextual trust softly selects among multiple propagation operators for each relation type via a learnable context-conditioned selector.
3. Component-Decoupled Temporal Trust Memory: Independent temporal memory states with non-uniform decay rates are maintained for each evidence channel, allowing asynchronous trust evolution modeling (slow entity accumulation, mixed behavior, volatile context) and preventing mutual overwriting.
4. Uncertainty-Aware Prediction and Calibration: The model predicts trust probability and estimates prediction uncertainty, aligning confidence calibration (Brier Score, ECE, NLL) with empirical error risk for reliable downstream filtering.

   Figure 2: TCHG architecture with tri-trust evidence encoding, conditioned propagation, component-specific temporal memory, and uncertainty-aware calibration.

Empirical Evaluation and Numerical Results

TCHG was benchmarked against a suite of baselines (Linear, Guardian, TrustGNN, CAT, HAN, RGCN, HGT) on three public datasets (Epinions, Ciao, CiaoDVD), representing both timestamped dynamic trust prediction and all-users random trust-link splits. Strong numerical improvements are observed:

• On Epinions (80%-10%-10%): TCHG achieves MRR 0.8811 (observed-user), 0.7018 (unobserved-user) versus CAT at 0.6758 and 0.4282, respectively.
• On CiaoDVD: MRR improvements from 0.8608 (CAT) to 0.9325 (TCHG).
• Probability Calibration: ECE reduced to 0.0158 (from >0.2647 without calibration), Brier score 0.0272, and selective prediction shows uncertainty scores reliably track empirical error.

  Figure 3: MRR comparison on Ciao under all-users random trust-link prediction splits for two split ratios.

  

  Figure 4: MRR comparison on CiaoDVD showing gains over baselines across splits.

Evidence-Controlled Propagation and Memory Analysis

A key claim supported by ablation is that tri-trust evidence must be used for propagation control—not as mere feature enrichment. Feature Injection Only and Attention Only variants result in significantly lower MRR. Ablations removing entity admission, behavior modulation, or context-operator selection confirm that all three components are necessary for top-ranked trust-link retrieval.

Component-wise temporal memory, with non-uniform decay and uncertainty, further improves test performance and reliability over memoryless or uniform-decay alternatives, especially for medium-to-long time gaps and observed-user scenarios. Memory statistics reveal entity memory as high-norm/low-uncertainty, contextual memory as low-norm/high-uncertainty, consistent with functional roles.

Figure 5: Selective prediction results—rejecting high-uncertainty edges yields higher sampled MRR, reflecting effective uncertainty calibration.

Figure 6: Sensitivity of overall MRR to the number of candidate heterogeneous propagation operators $K$.

Figure 7: Sensitivity of overall MRR to component-specific temporal decay assignments.

Figure 8: Entity, behavior, and contextual memory state statistics showing averaged norms, update norms, and uncertainty.

Robustness to Poisoning Attacks

Robustness evaluations under CAT-style trust-oriented and GNN-oriented poisoning attacks demonstrate that TCHG maintains superior absolute attacked MRR for both observed-user and unobserved-user cases, even as relative degradation varies with attack type and split. The conditioned propagation and component memory modules impart additional resilience to local trust manipulation and contextual graph perturbations.

Figure 9: Robustness comparison under poisoning attacks, showing absolute and relative performance for TCHG and CAT across splits and task types.

Practical and Theoretical Implications

TCHG offers a principled mechanism for dynamic trust modeling both in timestamped and timestamp-free settings. By disentangling heterogeneous evidence and controlling propagation functionally, TCHG achieves reliable trust prediction with calibrated probabilities suitable for downstream risk-sensitive applications (filtering, recommendation, manipulation detection). The evidence-controlled approach generalizes to other relational prediction tasks where channel evidence plays asymmetric roles in information flow.

Algorithmically, TCHG demonstrates effective integration of evidence abstraction, dynamic propagation, asynchronous memory evolution, and uncertainty calibration within a unified heterogeneous graph architecture. The observed gains are not attributable to model scale or feature stacking, but to propagation-level functional control.

From an AI development perspective, TCHG previews a methodological trajectory for future trust-aware and evidence-modulated graph learning, especially in domains where dynamic context and entity behavior are critical.

Conclusion

TCHG advances trust prediction by decomposing evidence and integrating tri-trust conditioned propagation, component-decoupled temporal memory, and uncertainty calibration. It surpasses prior baselines in both ranking and probabilistic reliability, with strong robustness properties. The approach sets a precedent for channel-aware information control in dynamic heterogeneous graphs and opens pathways for rich multimodal trust modeling and adaptive robustness defense.