Relational Learning Systems as Collaborators with Human Experts in Knowledge Discovery

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

Machine learning systems are proving to be useful tools for knowledge discovery. Their abilities to process large amounts of data and to consider large hypothesis spaces complement the abilities of human experts to analyze in great depth a few data points and a few potential hypotheses. Nevertheless, except for the initial presentation of data and the final presentation of a hypothesis, current machine learning systems and human experts do not interact. The goal of this project is to explore ways to enable machine learning system and human experts to act more as a team, taking advantage of both the computer's speed and the human's knowledge and skills. The PI will study the issues involved in building and using collaborative machine learning systems, through the development and testing of a system with capabilities to do the following: 1) Maintain, cluster, and summarize alternative hypotheses that explain the data, rather than providing a single answer based on a general-purpose heuristic; 2) Propose to human experts practical sequences of experiments to refine or distinguish between competing hypotheses, and justify these experiments to the experts; 3) Provide non-numerical justification for hypotheses, such as relating them to prior beliefs or illustrative examples (in addition to providing numerical accuracy estimates); 4) Answer an expert's questions regarding alternative hypotheses; and 5) Consult the expert regarding anomalies or surprises in the data. For focus, the developed system will be a relational learning system. Such systems already have advantages for use as collaborators in knowledge discovery, because they can utilize declarative background knowledge provided by an expert, and return hypotheses in the form of human-comprehensible rules. For further focus, the project will apply collaborative learning systems to a family of challenging scientific discovery tasks within the general area of drug design. This research should impact the areas of knowledge discovery, machine learning, and data mining in bringing about a shift toward the development of systems that can be tailored by different users to application areas as diverse as genomics, logistics, education and economic forecasting, and which can then interact with human experts in these areas to make discoveries neither could have made alone. More specifically, a collaborative machine learning system called COLLEAGUE will be made publicly available as a research testbed for use on tasks where large or complex data sets must be analyzed and human knowledge or skills are required for success.