BIGDATA: F: New Algorithms of Online Machine Learning for Big Data

该项目正在开发创新的、理论上严格的算法，以学习连续到达的(流)数据。解决的具体挑战是类别不平衡(要学习的概念之一非常罕见，如在疾病检测中)、对获取特征的成本限制(例如，计算昂贵的图像处理)、以及对获得类别标签的成本限制(例如，人工注释)。由于数据分布严重不平衡、超高维特征、大量标签、高度复杂的约束等各个方面的复杂性增加，该项目开发的算法可以有效地应对流数据中的大数据挑战。该项目还将有助于培训未来的大数据分析专业人员，包括参与爱荷华大学的本科生暑期研究计划和高中生培训计划。大多数致力于在线学习算法及其分析的工作都是以最小化对称度量(例如分类错误)为目标的，而没有考虑大数据中出现的实际约束。这个项目通过开发在线学习算法来解决不平衡数据，以最小化不对称度量，包括F-Score、ROC曲线下面积以及精度和召回率曲线下面积。很好地逼近这些非对称度量的凸或非凸代理损失函数被构造并以在线方式最小化。该项目还通过探索随机化算法、主动学习和凸优化技术，开发了大数据环境中出现的三种约束条件下的在线算法，即计算成本约束、查询成本约束和复杂不等式约束。开发的算法正在生物医学语义索引、社交媒体挖掘和图像标注等实际应用中进行评估。

This project is developing innovative, theoretically rigorous algorithms to learn from continuously arriving (streaming) data. Specific challenges addressed are class imbalance (one of the concepts to be learned is very rare, as in disease detection), cost constraints on both obtaining features (e.g., computationally expensive image processing), and cost constraints on obtaining class labels (e.g., human annotation.) The algorithms developed in this project make it possible to effectively address big data challenges in streaming data due to increased complexities in various aspects such as heavily imbalanced data distributions, ultrahigh dimensional features, a large number of labels, highly complex constraints, etc. The project will also contribute to training future professionals in big data analytics, including participation in the University of Iowa's undergraduate summer research program and high school student training program.Most work devoted to online learning algorithms and their analysis were developed with the goal of minimizing a symmetric measure (e.g., the classification error) and without considering practical constraints arising in big data. This project addresses imbalanced data by developing online learning algorithms for minimizing asymmetric measures including F-score, area under the ROC curve, and area under precision and recall curve. Convex or non-convex surrogate loss functions that well-approximate these asymmetric measures are constructed and minimized in an online fashion. The project also develops online algorithms under three types of constraints arising in big data context namely constraints on computing costs, on query costs, and complex inequality constraints, by exploring techniques in randomized algorithms, active learning and convex optimization. The developed algorithms are being evaluated in real applications including biomedical semantic indexing, social media mining, and image annotation.