Using norms to guide and coordinate interactions has gained tremendous attention in the multi-agent community. However, new challenges arise as the interest moves towards dynamic socio-technical systems, where human and software agents interact, and interactions are required to adapt to human's changing needs. For instance, different agents (human or software) might not have the same understanding of what it means to violate a norm (e.g., what characterizes hate speech), or their understanding of a norm might change over time (e.g., what constitutes an acceptable response time). The challenge is to address these issues by learning the meaning of a norm violation from limited interaction data. For this, we use batch and incremental learning to train an ensemble of classifiers. Ensemble learning and data-sampling handle the imbalanced class distribution of the interaction stream. At the same time, the training approaches use different strategies to ensure that the ensemble models reflect the latest community view on the meaning of norm violation. Batch learning uses weight assignment, while incremental learning continuously updates the ensemble models as community members interact. Here, we extend our previous work by creating a different balance strategy for online learning and integrating interpretability to understand norm violations. Additionally, we evaluate the proposed approaches in the context of Wikipedia article edits, where interactions revolve around editing articles, and the norm in question is prohibiting vandalism. Lastly, we conduct ablation studies to compare the ensemble's performance against a single model approach and to examine the behavior of two data sampling techniques. Results indicate that the different machine learning frameworks can learn the meaning of a norm violation in a setting with data imbalance and concept drift, although with significant differences.

使用规范来指导和协调交互已经在多智能体社区中引起了极大的关注。然而，随着人们的兴趣转向动态社会技术系统，人类和软件代理相互作用，并且需要相互作用来适应人类不断变化的需求，新的挑战出现了。例如，不同的主体（人或软件）可能对违反规范的含义（例如，仇恨言论的特征）有不同的理解，或者他们对规范的理解可能会随着时间的推移而改变（例如，什么构成了可接受的规范）。响应时间）。挑战在于通过从有限的交互数据中了解违反规范的含义来解决这些问题。为此，我们使用批量和增量学习来训练分类器集合。集成学习和数据采样处理交互流的不平衡类别分布。同时，训练方法使用不同的策略来确保集成模型反映社区对规范违反含义的最新看法。批量学习使用权重分配，而增量学习随着社区成员的互动不断更新集成模型。在这里，我们通过为在线学习创建不同的平衡策略并整合可解释性来理解规范违规行为来扩展我们之前的工作。此外，我们在维基百科文章编辑的背景下评估所提出的方法，其中交互围绕编辑文章进行，而所讨论的规范是禁止破坏行为。最后，我们进行消融研究，将集成的性能与单一模型方法进行比较，并检查两种数据采样技术的行为。结果表明，不同的机器学习框架可以在数据不平衡和概念漂移的环境中学习违反规范的含义，尽管存在显着差异。