SGER: Exploiting Contextual Knowledge to Design Input Representations for Machine Learning

Oregon State University, Corvallis OR

Investigators

Abstract

Non-specialists often have difficulty designing the features used to formulate machine learning problems. This small grant for exploratory research will explore the possibility of automating this "feature engineering" aspect of machine learning problems. Specifically, this project will investigate the feasibility of representing the background knowledge in a modern knowledge representation language and then automatically (or semi-automatically) deriving the relevant input features. The project will study two application problems where there is extensive background knowledge but sparse training data: (a) predicting future grasshopper infestations; and (b) understanding the spread of West Nile virus. This research will provide a deeper understanding of the process of feature engineering and help articulate an agenda for future research.