III: Small: Collaborative Learning with Incomplete and Noisy Knowledge

大数据的加速发展为知识发现创造了大量宏观层面的信息。但在微观层面上，我们只能对大多数个人用户进行少量观察。这阻碍了对不同用户之间的细微模式和异质性的探索，从而以每个用户为基础提高大数据分析的效用。该项目的目标是开发一套算法解决方案，以协作方式执行在线学习，其中个性化学习解决方案积极与用户交互以获取反馈，并相互协作以从不完整和嘈杂的输入中学习。该项目扩大了统计学习在许多重要领域的效用，例如医疗保健、商业智能、众包和网络物理系统，其中自动化决策模型建立在多样化、嘈杂和异构监督的基础上。这些研究活动将纳入资料检索、机器学习和数据挖掘领域的学生培训和教育的教材。该项目包括三个协同研究重点。首先，它开发了一系列上下文强盗算法来对网络用户进行协作在线学习。估计和利用用户之间的依赖关系来协作更新个性化的强盗参数。其次，它开发了原则性的解决方案来优化在线学习的特定任务和一般损失函数，这使得协作学习解决方案能够达到更重要的现实应用，如信息检索和用户行为建模。第三，它对反馈来源的可靠性进行建模和区分，以优化整体在线学习效果，这在健康信息学、众包和网络物理系统等应用中尤为重要。项目的预期成果包括：1)开发的在线学习解决方案的开源实现；2)评价语料库，使研究者能够在相关领域进行后续研究。

The accelerated growth of Big Data has created enormous amount of information at the macro level for knowledge discovery. But at the micro level, one can only expect a handful of observations in most individual users. This hinders the exploration of subtle patterns and heterogeneities among distinct users for improving the utility of Big Data analytics at a per-user basis. The objective of this project is to develop a set of algorithmic solutions to perform online learning in a collaborative fashion, where personalized learning solutions actively interact with users for feedback acquisition and collaborate with each other to learn from incomplete and noisy input. This project amplifies the utility of statistical learning in many important fields, such as healthcare, business intelligence, crowdsourcing, and cyber physical systems, where automated decision models are built on diverse, noisy and heterogeneous supervision. The research activities will be incorporated into teaching materials for student training and education in the areas of information retrieval, machine learning and data mining. This project consists of three synergistic research thrusts. First, it develops a family of contextual bandit algorithms to perform collaborative online learning over networked users. Dependency among users is estimated and exploited to collaboratively update the individualized bandit parameters. Second, it develops principled solutions to optimize task-specific and general loss functions for online learning, which enables the collaborative learning solutions reach more important real-world applications, such as information retrieval and user behavior modeling. Third, it models and differentiates the reliability of the sources of feedback to optimize the overall online learning effectiveness, which is especially important in the applications such as health informatics, crowdsourcing and cyber physical systems. Expected outcomes of the project include: 1) open source implementations for the developed online learning solutions; and 2) evaluation corpora that will enable researchers to conduct follow-up research in related domains.