III: Small: Robust and Scalable Reputation Management and Recommender Systems Using Belief Propagation

声誉和推荐系统在在线营销、Web服务、P2P计算、电子商务、社交环境和教育中有着广泛的应用。本研究的目标是开发可靠的，可扩展的和可靠的计划，也是弹性的恶意行为。该方法是基于查看信誉和推荐系统作为解决边际概率分布函数从复杂的全球（联合分布）功能的许多变量（用户，项目或服务提供商，和评级）。这些边际概率分布是表示信誉值（在信誉系统中）和要向用户预测的评级（在推荐系统中）的变量的函数。然而，计算这些边际分布在计算上是禁止的（即，与变量的数量呈指数关系）。因此，本研究使用因子图或成对马尔可夫随机场表示信誉和推荐系统，并利用置信传播（BP）算法有效地（在线性复杂度）解决这些边缘分布。具体而言，该项目包括以下研究：（1）研究基于BP的信誉管理和推荐系统在各种图形模型上的一般理论，并开发新的算法;（2）通过数学分析和密集的模拟研究所开发算法的收敛性，可扩展性和鲁棒性;（3）开发基于信念传播的迭代信任和信誉管理（BP-ITRM）系统，并使用真实数据集和进行用户研究将其与当前技术状态进行比较;以及（4）自适应地学习针对信誉和推荐系统的各种攻击策略，确定这些攻击的影响，并且减小它们的影响。 该项目预计将通过开发一个新的推荐系统的声誉管理框架，并提供有效的方法来处理信息过载和访问相关信息的算法，在理论和实践上做出贡献。预计这项工作将推动有效的在线产品和信息服务的技术。这项研究所产生的技术将在许多领域带来广泛的好处，包括在线服务，P2P和分布式计算系统，电子商务，商业，社会环境，教育，国家安全和经济。研究成果有望对计算机科学、信息论和统计推断等领域做出理论贡献。该项目提供了一个独特的机会，培养研究生，并让本科生接触不同领域（计算机科学，信息论，统计推断）的交叉研究。项目网站（http：//www.ece.gatech.edu/research/labs/WCCL/Security6.html）用于向广大研究人员、学生和行业从业人员传播由此产生的出版物、数据集（从用户研究和数学模型中获得）和课程材料。

Reputation and recommender systems have widespread use in online marketing, web services, P2P computing, e-commerce, social settings, and education. The objective of this research is to develop reliable, scalable and dependable schemes that are also resilient to malicious behaviors. The approach is based on viewing both reputation and recommender systems as solving for marginal probability distribution functions from complicated global (joint-distribution) functions of many variables (users, items or service providers, and ratings). These marginal probability distributions are functions of the variables representing the reputation values (in reputation systems) and the ratings to be predicted to the users (in recommender systems). However, computing these marginal distributions are computationally prohibitive (i.e., exponential with the number of variables) for large scale reputation and recommender systems. Therefore, this research represents the reputation and recommender systems using factor graphs or Pairwise Markov Random Fields, and utilizes the Belief Propagation (BP) algorithm to efficiently (in linear complexity) solve for these marginal distributions. In particular, the project includes research to: (1) study the general theories of BP-based reputation management and recommender systems on various graphical models and develop novel algorithms; (2) study the convergence, scalability, and robustness of the developed algorithms via mathematical analysis and intensive simulations; (3) develop a Belief Propagation based Iterative Trust and Reputation Management (BP-ITRM) system and compare it with the current state of the art using real-life datasets and conducting user studies; and (4) adaptively learn various attack strategies against the reputation and recommender systems, determine the impacts of such attacks, and decrease their impact. The project is expected to make contributions to both theory and practice by developing a new reputation management framework for recommender systems, and algorithms that provide effective ways to deal with information overload and access to relevant information. It is anticipated that the work will drive the technology for effective online products and information services. Technologies resulting from this research will bring a broad range of benefits in many areas including online services, P2P and distributed computing systems, e-commerce, business, social settings, education, national security and the economy. The research results are expected to make theoretical contributions relevant to computer science, information theory and statistical inference. This project offers a unique opportunity to train graduate students and expose undergraduate students to cross-cutting research in different fields (computer science, information theory, statistical inference). The project website (http://www.ece.gatech.edu/research/labs/WCCL/Security6.html) is used to disseminate resulting publications, datasets (obtained from user studies and mathematical models), and course materials to broad communities of researchers, students and industry practitioners.