Trust and reputation management underpins reliable interactions in distributed networks, yet existing trust models rely solely on forward propagation of interaction-based trust signals. They lack robust mechanisms to enforce accountability for the propagated trust signals when negative interactions occur. In addition, such models often fail to initialize newly joined nodes with sparse interaction history, leading to the cold-start problem. In this paper, we propose RepuLink, a two-layer reputation model that couples an endorsement network with an interaction feedback network. RepuLink integrates two concurrent backward propagation mechanisms: Backward Endorsement Penalty Propagation (BEPP), which recursively penalizes endorsers of misbehaving nodes, and Backward Endorsement Reward Propagation (BERP), which rewards endorsers of well-performing nodes. Together, RepuLink enforces endorsement accountability and incentivizes positive behaviors, which form a positive interaction feedback loop. The endorsement layer further provides explainable, endorser-weighted trust initialization for newly joined nodes. Experiments on real-world datasets against representative trust propagation baselines demonstrate that RepuLink outperforms across four evaluation metrics in both interaction-only and full two-layer settings, while preserving comparable efficiency.

翻译：[translated abstract in Chinese] 信任与声誉管理是分布式网络中可靠交互的基础，然而现有信任模型仅依赖于基于交互的信任信号的前向传播。当负面交互发生时，这些模型缺乏强有力的机制来对传播的信任信号进行问责。此外，此类模型通常难以对交互历史稀疏的新加入节点进行初始化，从而导致冷启动问题。本文提出RepuLink，一种双层声誉模型，它将背书网络与交互反馈网络相结合。RepuLink集成了两种并发的反向传播机制：反向背书惩罚传播（BEPP），用于递归惩罚不当行为节点的背书者；以及反向背书奖励传播（BERP），用于奖励表现良好节点的背书者。通过协同作用，RepuLink强化了背书问责制并激励积极行为，从而形成正向交互反馈循环。背书层还能为新加入节点提供可解释的、基于背书者权重的信任初始化。在真实数据集上针对代表性信任传播基线模型的实验表明，RepuLink在仅交互层与完整双层的设置下，均能在四项评估指标上取得更优性能，同时保持相当的效率。