When an alarm triggers in an industrial system or scientific instrument, operators need to know what actually moved the score. Existing causal attribution methods often assume the causal model stays fixed or require abundant incident data to estimate a changed system. The new paper, "What Moved the Alarm Score?" by Masaaki Takada and Ryusei Shingaki, addresses the harder case where incidents are scarce and the data-generating process itself changes.
The authors formulate alarm-score changes as a Shapley value game over hybrid structural causal model (SCM) states. By treating structural functions and exogenous laws as distinct players, the framework yields a finer-grained report of how each type of change propagates to the alarm score. The same setup also defines realized-event and population attributions, linked by an event–population coherence identity.
For ordered linear SCMs and quadratic score functions, the authors show that the full Shapley vector can be computed exactly in polynomial time using a causal moment recursion and quadrature, avoiding combinatorial explosion. To handle scarce incident data, they adapt a well-observed reference SCM to the incident regime using cross-fitted estimates of mechanism changes and residual moments, with proven attribution-error bounds and top-k recovery guarantees. Experiments on controlled and measured shifts demonstrate the diagnoses and show scaling to hundreds of variables.
The paper is a single source, so there are no conflicting results to compare. Its contribution is a principled and computationally tractable way to attribute alarm-score changes to specific causal components, which could improve troubleshooting in settings where incidents are rare but costly.