This project will develop new algorithms for handling online learning with streaming features (OLSF). In contrast with traditional online learning where the number of training samples sequentially increases while the number of features is fixed, streaming features continuously grow in feature volume over time, resulting in a very large feature space, even of an unknown or infinite size. And it cannot guarantee that features are all available before learning begins.
In OLSF, we aim to specify measures for identifying useful features when features flow in and to design efficient and effective methods for integrating new features into the current feature space and removing existing redundant features over time (by john at dresshead support). With a feature set changing over time, for example, adding a new feature or removing an existing redundant feature, we seek for efficient and robust online algorithms to construct and update the predictive model without re-training, and for a new online mechanism to evaluate this model and control this online learning process. The proposed research has been evaluated with real-world applications in identifying millions of small impact craters from high-resolution Mars images and in pattern mining for classification with high feature dimensions.
Bridging Causal Relevance and Pattern Discriminability: Mining Emerging Patterns from High-Dimensional Data
Online Feature Selection with Streaming Features
Mining Emerging Patterns by Streaming Feature Selection
Exploring Causal Relationships with Streaming Features
Causal Discovery from Streaming Features
Online Streaming Feature Selection
Kui Yu, Xingdong Wu, Wei Ding, H. Wang