Exploiting Evidence-dependent Sensitivity Bounds
This work addresses a computational bottleneck for researchers and practitioners in probabilistic modeling, offering incremental improvements to sensitivity analysis methods.
The paper tackles the problem of efficiently identifying parameters with large effects on output probabilities in complex networks by studying evidence-dependent sensitivity bounds, demonstrating that these properties allow for establishing upper bounds on sensitivity values and identifying regions where low-sensitivity parameters can still have significant impacts.
Studying the effects of one-way variation of any number of parameters on any number of output probabilities quickly becomes infeasible in practice, especially if various evidence profiles are to be taken into consideration. To provide for identifying the parameters that have a potentially large effect prior to actually performing the analysis, we need properties of sensitivity functions that are independent of the network under study, of the available evidence, or of both. In this paper, we study properties that depend upon just the probability of the entered evidence. We demonstrate that these properties provide for establishing an upper bound on the sensitivity value for a parameter; they further provide for establishing the region in which the vertex of the sensitivity function resides, thereby serving to identify parameters with a low sensitivity value that may still have a large impact on the probability of interest for relatively small parameter variations.