KNOWLEDGE DISCOVERY BY DESIGN

Traditional work in knowledge discovery focuses on identifying statistical associations. However, increasing numbers of applications require knowledge about cause and effect, rather than mere association. For example, while a decision tree can estimate the probability that a patient will contract a particular disease, it does not provide guidance about how to prevent that disease. Similarly, while associational rules can predict which products will be purchased together, they do not indicate what actions will increase sales of those products. In this talk, I describe some surprising and powerful new opportunities for discovering causal knowledge from observational data. These opportunities exploit recent developments in relational and temporal knowledge discovery. They also draw on a large set of manual techniques for knowledge discovery called quasi-experimental designs. These techniques are widely employed by economists, medical researchers, and social scientists, yet they have never been exploited within knowledge discovery algorithms. Together, these opportunities point to a broad class of algorithms for discovering causal knowledge that can be developed in the next several years.