CIF: Small: Online Algorithms for Streaming Structured Big-Data Mining

点击翻译按钮获取中文摘要

In today's big data age, a lot of streaming big-data is generated around us. Most of this is either not stored or stored only for short periods of time, e.g., streaming videos or changing social network connections. This project develops novel online algorithms for dimensionality reduction and structural information recovery from incomplete or distorted data.This research makes several contributions to both the theory and the practice of structure recovery from streaming big-data. (1) The investigator and her team are developing provably correct online algorithms for robust structure (subspace or support) tracking from undersampled, outlier-corrupted or otherwise highly noisy streaming data. While batch approaches for these problems have been well-studied, the online problem is largely open. Our theoretical results are among the first correctness results for the robust subspace tracking (online robust PCA) and robust support tracking problems. Online algorithms are useful because they are faster and need lesser storage compared to most batch techniques. Moreover, our online algorithms remove a key limitation of batch approaches by exploiting certain extra temporal dependency assumptions: they allow significantly more correlated support change compared with batch methods. (2) The investigator is also developing novel provably accurate solutions for the online robust sparse-PCA problem, as well as several other related problems. (3) Finally, this project is helping to produce a well-trained and diverse future workforce. We expect our solutions to significantly transform the state-of-the-art in various big-data analytics applications, e.g., streaming video analytics, mobile video chats, autonomous vehicle or airplane navigation in foggy or rainy environments, anomalous or suspicious behavior detection from dynamic social network connectivity data.