Tractable Bounding of Counterfactual Queries by Knowledge Compilation
summary
Claims that compiling a causal model once into a symbolic compilation (a single parameterized arithmetic circuit) lets EM-based bounding of partially identifiable counterfactuals reuse the circuit across many parameter settings, cutting repeated compilation cost. Implementation adds parallelization across EM initializations and per-record posterior queries, and experiments report roughly 2x speedup from parallel runs, ~5x from parameterized compilation, and overall up to an order of magnitude faster than repeated Bayesian network inference.
AI-written summary · about