CAREER: Learning- and Incentives-Based Techniques for Aggregating Community-Generated Data

互联网带来了大量用户偏好、行为和信仰方面的新数据来源。研究人员和工程师都渴望汇总和解释这些数据。然而，网站有时不能激励高质量的贡献，导致数据质量不稳定。此外，传统学习理论所做的假设在这种情况下也会失效。该项目旨在创建基本的机器学习模型和算法，以解决和解释在大型社区聚集当地信仰时出现的问题，并促进对如何激励高质量贡献的最先进理解。研究可分为三个方向：1.研究方向：开发从社区标记数据中学习的数学基础和算法。这个方向包括为来自不同来源（可能是自利的或恶意的）的数据开发学习模型，并利用这些模型的洞察力来设计有效的学习算法。理解和设计更好的激励机制。这个方向包括建模众包贡献，以确定在系统中包含哪些特性以鼓励最高质量的贡献。引入新颖的经济驱动的意见聚合机制。这包括形式化预测市场应该满足的属性，并利用机器学习和优化的思想来推导满足这些属性的可处理的市场机制。这项研究将对行业产生明显的影响，特别是对基于网络的众包。PI将通过参与研讨会和指导项目来实现她的长期目标，即吸引和留住计算机科学领域的女性。结果将在http://www.cs.ucla.edu/~jenn/projects/CAREER.html上发布。

The Internet has led to the availability of novel sources of data on the preferences, behaviors, and beliefs of massive communities of users. Both researchers and engineers are eager to aggregate and interpret this data. However, websites sometimes fail to incentivize high-quality contributions, leading to variable quality data. Furthermore, assumptions made by traditional theories of learning break down in these settings.This project seeks to create foundational machine learning models and algorithms to address and explain the issues that arise when aggregating local beliefs across large communities, and to advance the state-of-the-art understanding of how to motivate high quality contributions. The research can be split into three directions:1. Developing mathematical foundations and algorithms for learning from community-labeled data. This direction involves developing learning models for data from disparate (potentially self-interested ormalicious) sources and using insight from these models to design efficient learning algorithms.2. Understanding and designing better incentives for crowdsourcing. This direction involves modeling crowdsourcing contributions to determine which features to include in systems to encourage the highest quality contributions.3. Introducing novel economically-motivated mechanisms for opinion aggregation. This involves formalizing the properties a prediction market should satisfy and making use of ideas from machine learning and optimization to derive tractable market mechanisms satisfying these properties.This research will have clear impact on industry, especially for web-based crowdsourcing. The PI will pursue her long-term goal of attracting and retaining women in computer science via her involvement in workshops and mentoring programs. Results will be disseminated at http://www.cs.ucla.edu/~jenn/projects/CAREER.html.