Zoom Meeting: https://umass-amherst.zoom.us/j/93178231162?pwd=WkpXclpKaEJLQjBYUkVKNTZweXpPUT09
Abstract
Causal modeling is central to many areas of artificial intelligence, including complex reasoning, planning, knowledge-base construction, robotics, explanation, and fairness. An active community of machine learning researchers develops and enhances algorithms that learn causal models from data, and this work has produced a series of impressive technical advances. However, evaluation techniques for causal modeling algorithms have remained somewhat primitive, limiting what we can learn from the experimental studies of algorithm performance, constraining the types of algorithms and model representations that researchers consider, and creating a gap between theory and practice. We argue for expanding the standard techniques for evaluating algorithms that construct causal models. Specifically, we argue for the addition of evaluation techniques that examine interventional measures rather than structural or observational measures, and that evaluate those measures on empirical data rather than synthetic data. We survey the current practice in evaluation and show that the evaluation techniques we advocate are rarely used in practice. We show that the techniques we advocate are feasible, that sufficient data sets are currently available to conduct such evaluations, and that these techniques produce substantially different results than using structural measures and synthetic data. We also provide a protocol for generating observational-style datasets from experimental data, allowing the creation of a large number of data sets suitable for evaluation of causal modeling algorithms. Using these data sets, we perform an evaluation of current causal modeling algorithms, using a significantly larger collection of data sets than was previously available.
Advisor: David Jensen