Modeling Trust in Open, Dynamic Multi-agent Systems and Developing Framework for Predicting Consumer-generated Reviews' Helpfulness

这个建议的研究计划的目标是双重的：（a）构建一个有效的信任模型，开放的，动态的多代理系统，和（B）开发一个框架建模的有用的在线评论。 信任建模对于开放的多Agent系统至关重要，在开放的多Agent系统中，Agent需要找到可信任的伙伴来帮助完成任务，而环境中可能存在欺骗性Agent。我们计划建立一个信任模型，该模型不仅在代理可以收集足够的证据来评估他人的可信度时有效，而且在信任证据不可用时也有效（例如，在新代理不断加入和离开的高度动态系统中）。我们的方法是计算代理的信任作为直接和间接证据的准确函数，并利用激励机制和机器学习来促进诚实和学习新的和现有的代理之间的相似性和差异。我们期望能够得到一个对电子市场、社交网络、车载自组织网络、云计算环境和对等系统等重要应用领域都有价值的信任模型。 在线产品评论已成为帮助消费者做出良好购买决策的主要信息来源。然而，由于大量的可用评论以及它们在有用程度上的巨大差异，消费者确实需要能够发现并向他们推荐最有用评论的系统。我们想开发一个严格的框架来推断评论的有用性，它使用概率方法来制定有用性推断作为一个优化问题。我们将使用我们的帮助预测初步模型作为起点，但将大大扩展它，以获得一个更复杂的模型，显着提高性能和准确性。这个框架预计将是有用的几个应用程序：搜索引擎，电子商务网站，在线社区和推荐系统可以利用这个框架提供最有帮助的评论，合适的产品，服务和/或供应商给他们的用户。

The goal of this proposed research program is twofold: (a) to construct an effective trust model for open, dynamic multi-agent systems, and (b) to develop a framework for modeling the helpfulness of online reviews. Modeling trust is of vital importance to open multi-agent systems where agents need to find trustworthy partners to help fulfill their tasks while deceptive agents may exist in the environment. We plan to build a trust model that works effectively not only when agents can gather sufficient evidence to evaluate the trustworthiness of others, but also when trust evidence is unavailable (e.g., in highly dynamic systems with new agents continually joining and leaving). Our approach is to compute an agent's trust as an accurate function of direct and indirect evidence, and to make use of incentive mechanisms and machine learning to promote honesty and learn the similarities and dissimilarities between new and existing agents. We expect to achieve a trust model that is valuable to many important application domains such as electronic marketplaces, social networks, vehicular ad-hoc networks, cloud computing environments, and peer-to-peer systems, etc. Online product reviews have become a major source of information to help consumers make good purchase decisions. However, due to the vast number of available reviews and their great difference in the level of helpfulness, consumers really need systems that can discover and recommend the most helpful reviews to them. We would like to develop a rigorous framework for inferring the helpfulness of reviews, which uses a probabilistic approach to formulate helpfulness inference as an optimization problem. We will use our preliminary model for helpfulness prediction as a starting point, but will extend it considerably to obtain a more complex model with significantly increased performance and improved accuracy. This framework is expected to be useful for several applications: Search engines, e-commerce websites, online communities and recommender systems can utilize this framework to offer the most helpful reviews, suitable products, services and/or vendors to their users.