Answers are grounded in this paper and cite their sources.
In high-stakes domains, a discovered relationship is useful only if its holding conditions can be stated explicitly and inspected by an expert. Causal-discovery algorithms return a graph of adjacent variables, but none states whether a parent is necessary, sufficient, or merely contributory, and post hoc explanations need not match the model’s actual computation. We present Bipolar Logic Networks (BLN), a neural-symbolic architecture with two interpretable paths—a sufficient disjunctive-normal-form path and a necessary conjunctive-normal-form path—fused by a learned gate; every reported configuration is a trained minterm or clause exercised in the predictions it makes. Across the Sachs, ALARM, and CHILD networks, 162 synthetic datasets, and a four-environment intervention experiment, BLN recovered nearly the entire ground-truth skeleton (recall 0.95, 0.96, 1.00), and on Sachs 12 of 18 recovered true edges were missed by both PC and GES. Conjunctions validated as wholes on held-out data reached positive purity several-fold above base rate, and non-empty stability separated a stable core (Jaccard 0.79–0.88 over 12–13 conditions) from condition-specific targets. Within the boundary that outputs are mechanism candidates, not identified causal effects, BLN complements structure learners by supplying roles, validated conditions, and stability that graphs lack.
No takes yet. Share an insight, caveat, or question.
Li Shidong (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: