TWC: Small: Statistical Models for Opinion Spam Detection Leveraging Linguistic and Behavioral Cues

在线意见现在在决策中发挥着关键作用，并影响着我们生活的方方面面。选择去哪吃饭、住哪、上哪所大学、读哪本书、看哪位医生，甚至是投票给哪一位政治候选人，都在很大程度上受到众包意见的影响。然而，据估计，网站上高达30%的评论是虚假的。由于美国经济的很大一部分正受到社会舆论的驱动，这对公众构成了严重的风险（例如，通过误导投资于低质量的产品、服务或医生）。联邦贸易委员会的意见可能很快就会认为网上欺诈是非法的，是法律的罪行。检测虚假的在线意见是一个迫切的研究领域。否则，在线社交媒体可能会继续发展而不被发现。该项目旨在开发新的欺骗检测算法，以识别欺诈行为。它协同整合了计算语言学，行为建模和统计机器学习的技术，以推进这一领域的知识。该项目包括四个方面的研究工作：1）从大规模噪声人群数据和小规模领域专家编码数据中学习欺骗分类器的新方法，2）将评论者的“垃圾信息”视为具有观察到的行为足迹的潜在的无监督模型，3）用于联合建模评论、评论者及其利用固有强化关系的语言和行为模式的关系体系结构，以及4）混合来自所有方法的线索的总体评分机制和端到端验证框架。该项目开发的技术可以（1）减少电子商务中的营销，消费者和经济风险;（2）改善用户分析，检测在线骚扰，偏执，巨魔和其他与国家安全密切相关的社交媒体欺诈;（3）开发的过渡技术课程/教程，吸引代表性不足的学生，包括少数民族和妇女。其结果是一套新颖的，有原则的，可扩展的技术来过滤意见垃圾邮件在大规模。

Online opinions now play a pivotal role in decision making and influence a wide spectrum of our lives. Choices of restaurants at which to dine, places to stay, universities to attend, books to read, doctors to consult, and even political candidates to vote for, are largely influenced by crowdsourced opinions. However, it is estimated that up to 30% of reviews on websites are fake. As a larger part of the US economy is becoming driven by social opinions, it poses a serious risk to the general public (e.g., by getting mislead to invest on low quality products, services or doctors). The Federal Trade Commission Opinion may soon consider online fraud as unlawful and a legal offense. Detecting fake online opinions is an urgent research area. Otherwise, online social media might continue to progress undetected. This project aims to develop novel deception detection algorithms in order to identify fraudulent behavior. It synergistically integrates techniques from computational linguistics, behavioral modeling and statistical machine learning in order to advance knowledge in this area. The project consists of a four-pronged research effort: 1) novel methods to learn deception classifiers from large-scale noisy crowd data and small-scale domain expert coded data, 2) unsupervised models that treat "spamicity" of reviewers as latent with observed behavioral footprints, 3) a relational architecture for jointly modeling reviews, reviewers, and their linguistic and behavioral patterns leveraging inherent reinforcement relations, and 4) an ensemble scoring mechanism blending cues from of all approaches, and an end-to-end validation framework. The techniques developed in the project can (1) reduce the marketing, consumer, and economic risk in e-commerce; (2) improve user profiling, detecting online harassment, bigotry, trolls, and other social media fraud that are of major relevance to national security; and (3) transition techniques developed to courses/tutorials and attract underrepresented students, including minorities and women. The result is a suite of novel, principled, and scalable techniques to filter opinion spam at large scale.