AIS: Learning from Initially Labeled Nonstationary Streaming Data

点击翻译按钮获取中文摘要

The objective of this research is to develop a novel framework for analysis of large volumes of streaming data, where data characteristics change over time, and all new data are unlabeled and unstructured. The proposed work will be seminal for analyzing streaming nonstationary and unlabeled data while removing the unrealistic assumptions and simplifications made by existing approaches. The proposed approach uses small initial training data to label the currently unlabeled new data, creates an envelope around this data and shrinks the envelope to determine the core support region of the data. Samples are extracted from this region to serve as future training data to iteratively label the new incoming unlabeled drifting data. Once initialized, this approach never needs labeled data, and can indefinitely track the changes in data distribution. The primary intellectual merit is a new framework, addressing arguably one of the most challenging learning problems, accomplished by strategic integration of machine learning and computational geome-try. Significant fundamental knowledge will be obtained through formal development of this framework, which will then benefit many real-world applications, currently not properly addressed under existing ap-proaches. Broader Impacts: The proposed framework promises to bring us closer to truly adaptive and intelligent (brain-like) learning, and allow proper analysis of data drawn from aforementioned scenarios, whose ap-plications include network intrusion, cyber security, web-usage analysis, natural language processing, anomaly detection, climate change and energy demand analysis. Project's educational component will form Integrated Research and Learning Communities, drawing undergraduate students to whom this field has been mostly inaccessible.