EAGER: Harnessing the Power of Graph Data Analytics

在线社交网络中用户生成的内容的快速增长形成了大图数据的丰富来源。该项目将开发从这种图表数据中提取信息的有效技术。具体而言，EAGER项目将为以下任务开发高效算法：（1）机器学习和情感分析，如大词汇量会话语音识别，多标签学习和多极情感分析，以及（2）效应器检测和相关问题，如业务扩展和谣言阻止。本计画所发展之先进最佳化技术，将显著改善现有资料探勘演算法之效能，并促进有效近似技术之理论研究。此外，该项目将提供一个良好的平台，教育学生的实践经验。为了开发所描述的高效算法，该项目将利用非线性组合优化的最新发展，并将制定新的方法，使用离散凸分析的子模函数和对偶理论的离散DC（不同的凸）规划，以及反向采样技术的概率目标函数。这些新方法将丰富机器学习和情感分析的吞吐量。由于非线性组合优化仍处于早期阶段，有效的近似算法将在新的基础理论发展的基础上进行探索性和创新性的研究。将根据若干真实的世界数据集进行详细评价。

The rapid growth of user generated content in online social networks forms an abundant source of big graph data. This project will develop efficient techniques for extracting information from such graph data. Specifically, this EAGER project will develop efficient algorithms for the following tasks: (1) machine learning and sentiment analysis, such as large vocabulary conversational speech recognition, multi-label learning, and multi-polarity sentiment analysis, and (2) effector detection and related issues, such as business expansion and rumor blocking. The advanced optimization techniques developed in this project will significantly improve the performance of the existing data mining algorithms and the theoretical study of efficient approximation techniques. In addition, this project will provide an excellent platform for educating students with hands-on experience. To develop the efficient algorithms described, the project will take advantage of recent development in nonlinear combinatorial optimization, and will formulate new methodologies using discrete convex analysis for submodular functions and duality theory for the discrete DC (different of convex) programming, as well as reverse sampling techniques for probabilistic objective functions. These new methodologies will enrich the throughput of machine learning and sentiment analysis. Since nonlinear combinatorial optimization is still in its early stages, exploratory and innovative research on efficient approximation algorithm will be performed based on fresh fundamental theoretical developments. A detailed evaluation will be conducted based on several real world data sets.