Collaborative Anomaly Detection for Revenue Recovery Using Ensembles of Heterogeneous Data Agents

Abstract

Revenue management systems in digital platforms increasingly depend on fine-grained telemetry to ensure that contracted value is actually collected. Operational failures, instrumentation gaps, and adversarial behavior can introduce subtle discrepancies between reported and billable activity, giving rise to revenue leakage. Manual investigation and rule-based alarms often struggle to cover the breadth of heterogeneous data sources involved in modern billing pipelines. This paper examines collaborative anomaly detection for revenue recovery, in which multiple specialized data agents share signals to prioritize candidate losses. Each agent operates near a distinct telemetry stream, maintains its own detection model, and publishes anomaly scores that reflect local evidence of under-reported revenue. The central question is how to combine these partially overlapping, noisy views into actionable recommendations that align with downstream investigation capacity. The proposed framework models local anomaly scores as heterogeneous features, learns ensemble weights linked to historical recovery outcomes, and incorporates structural constraints derived from business rules. A linear decision layer maps aggregated scores to a ranking over candidate anomalies, while an optimization module selects subsets consistent with operational budgets. The study explores agent reliability modeling, cross-agent calibration, and robustness to missing or delayed signals in streaming settings. Empirical evaluation on synthetic and production-inspired datasets compares collaborative ensembles with isolated detectors and monolithic models trained on centralized logs. Under the examined scenarios, collaborative data agents allocate investigative effort toward events with higher estimated financial impact and expose anomalies that remain hidden to single-view methods.