Identifying Decision Effects through Instance-Level Causal Partitioning in Prescriptive Process Monitoring

Abstract: Prescriptive process monitoring seeks to improve process outcomes by recommending actions during ongoing cases. Choosing an action requires estimating how it would affect the outcome. This is challenging when cases reach the same decision through different execution paths and with different information available. We introduce instance-level causal graphs to represent case-specific causal structures and an instance-level causal-graph template to capture the structures compatible with an observed execution and its temporal order. Under stated causal assumptions, we show that the complete pre-decision history is a valid adjustment set in every graph represented by the template. Because a single case does not provide enough data to estimate a decision effect, we introduce causal decision variants (CDVs), which group cases that reach the same decision with the same set of observed pre-decision variables. We characterize when the decision effect remains identifiable within a CDV even if the underlying graphs differ and are unknown. Experiments show that the benefits of CDV-specific estimation depend on the setting. In a controlled synthetic benchmark, it improves conditional-effect estimation and ranking when decision effects vary across cases, while pooled models perform better when effects are constant. On semi-synthetic Sepsis data, only one CDV has sufficient support, and conditional-effect error is similar to the simplest pooled model. Neither benchmark shows a consistent advantage for population-average effect estimation.