Adaptive Online Mining of Big Data Streams

点击翻译按钮获取中文摘要

Big data streams are continuous flows of data that arrive in high volume and high velocity. Such data streams have become ubiquitous as many sources, such as sensor networks, business transactions and surveillance cameras, produce data continuously and rapidly. There has been a growing demand for real-time online analyzing and learning of big data streams so that patterns and models can be learned from such data in a timely manner to support fast decision making in a dynamic environment, for example, performing fraud detection at a point of sale. However, although machine learning techniques have become very effective in many applications, such as computer vision and speech recognition, learning a complex model (such as a deep neural network) from big data can be very time-consuming, making it impractical to work online. The objective of this research program is to accelerate machine learning programs and make it work in an online fashion. We will develop novel techniques for parallelizing machine learning models to speed up the learning process. We will address the following challenges in online learning of data streams. First, online algorithms are often constrained by space and time. Not all data can be stored. Algorithms often need to process the data in a single pass. However, many machine learning algorithms require a large number of passes over the data to find good solutions. How to adapt such learning algorithms to work online is an open challenge. Second, streaming data evolve over time with unknown dynamics, a phenomenon known as concept drift. Online learning from data streams should keep the model up to date and quickly adapt it to concept drift. Although methods have been proposed to train drift-adaptive online models, little has been done on adaptively learning complex nonlinear functions. Third, in streaming environments, data flow into the system at an unpredictable rate and the available resource may change dynamically due to resource sharing with other computing jobs. The processing system must keep up with the data rate and resource change. Resource adaptive online learning is highly needed. We will develop novel strategies for handling concept drift and resource constraints for learning complex functions online.  Parallel and distributed solutions over multiple processors or machines will be developed to speed up learning. Strategies that make trade-offs between resource consumption and the accuracy of a learned model according to resource conditions will be investigated. Anytime learning algorithms will be designed that can produce a best possible answer according to real-time constraints. Distributed online mining of big and fast data streams is still far from mature. The proposed research will advance the field by proposing novel solutions to its open challenges and will have a wide range of applications in various fields, e.g., fraud detection in real time and image recognition in a dynamic environment.