Aggregation Semantics for Prioritization in Cooperative Automated Decision-Making

Cooperative automated decision-making systems can combine predictive assessments with contextual knowledge to prioritize alternatives under resource constraints. The mechanism used to integrate these sources can influence whether contextual requirements are preserved or can be offset by predictive evidence. This study examines this issue in CADEMAS-ML, a recently proposed framework for cooperative automated decision-making that separates predictive and contextual decision pathways. Three aggregation regimes are analyzed: linear, weighted geometric, and minimum aggregation. The study combines theoretical analysis, controlled Monte Carlo experiments, and an employee-attrition case study. The theoretical analysis characterizes the conditions under which contextual vetoes are preserved during integration. The simulation study evaluates policy compliance, sensitivity to predictive uncertainty, ranking stability, and changes in the Top-K intervention set. The case study applies the same aggregation mechanisms to a reproducible CADEMAS-ML workflow based on predictive models and fuzzy contextual rules. The results show that aggregation semantics can affect policy compliance and prioritization outcomes. Under the considered conditions, zero-absorbing operators preserve contextual vetoes, whereas linear aggregation can allow predictive evidence to compensate for a contextual veto. This effect depends on the relative weighting of the two sources and on the characteristics of the decision population. Predictive uncertainty can also modify rankings and intervention sets, with its effects depending on the aggregation regime and intervention capacity. In the employee-attrition case study, changing only the aggregation mechanism produces different intervention sets from the same predictive and contextual assessments. These findings support treating aggregation as an explicit design choice in cooperative automated decision-making, since its semantics influence how contextual requirements are translated into automated prioritization.