III: Small: Integrating Casual Discovery and Feature Selection with Streaming Features

University Of Vermont & State Agricultural College, Burlington VT

Investigators

Abstract

With the advent of emerging massive datasets in image processing, biology, finance, and so on, traditional data mining systems face new challenges to induce knowledge and discover causal relations in dynamic streaming feature environments, where new features continuously stream in over time. These challenges include (1) continuous growth of feature volumes over time, (2) a huge feature space, even of unknown or infinite size, and (3) not all features being available before learning begins. These challenges call for a new learning paradigm with continuously increasing features. In this project, we take the increasing feature volumes as streaming features, and the corresponding learning problem is referred to as Online Learning with Streaming Features (OLSF). Since existing online learning efforts mostly deal with data with increasing observations but fixed feature dimensions, OLSF provides a unique chance to unfold and characterize pattern trends for dynamic systems with streaming features. This project aims to address two fundamental issues for OLSF: (1) causal discovery with sequentially increasing feature dimensions; and (2) causal relations for feature selection. We design novel methods and algorithms for causal discovery in OLSF and establish formal connections between casual discovery and feature selection by investigating the mutual benefits between them in the context of online stream feature learning. To evaluate the proposed research, we conduct empirical studies on a large body of benchmark datasets, as well as with a domain-specific real-world case study in personalized news filtering and summarization where the feature space changes over time. The new algorithms and techniques in this project will advance our ability to discover knowledge from dynamic systems using streaming features with bounded resources. The spectrum of the methods from the project will not only enrich our knowledge and understanding of pattern discovery and machine learning for dynamic systems, but also provide a new view to capture and characterize dynamic systems from a streaming feature perspective.