Reviews: On Tractable Computation of Expected Predictions

The manuscript considers basic statistical questions regarding reasoning about the expected outcome of a predictive model. Efficiently computing even the expectation (first moment) is a known challenge even for simple predictive models and simple generative models (e.g. The authors give a pair of generative and discriminative models (family of structured probabilistic circuits) that enables tractable computation of expectations (and higher order moments as well), in some cases approximately, b) provide algorithms for computing moments of predictions wrt generative models and c) show that the utility of the algorithms in handling missing data during prediction time compared to standard imputation techniques on some datasets. The paper is organized and written well, there are some good technical contributions. But I'm unable to get a good grasp on the overall significance and merit of this work - partly because the authors aren't convincing enough throughout the paper.