Exploiting Cross-KPI Confidence Patterns to Expand Automation Scope in 5G Network Management

Decisions to change configurations are often basedon predictions for Key Performance Indicators (KPIs). If thesepredictions are trustworthy, the decisions they support canbe automated. Confidence-based routing makes this separation,automating decisions backed by high-confidence predictions andescalating the rest for human review. Across three European5G production networks (500 to 12,000 cells, 41.8 million escalatedpredictions), we find that 86–89% of flagged predictionsdid not need escalation, overwhelming human reviewers at aratio exceeding 6:1. Confidence signals that feed the combinedscore cannot further separate genuine failures from spuriousescalations. However, cross-KPI confidence variance, the dispersionof scores across co-escalated KPIs at the same celland time, distinguishes uniform low confidence (cell-level weakpeers, predictions typically accurate) from uneven confidence(KPI-specific failure, 2.7–2.8× higher bound violation rates).This pattern holds across performance categories but invertsfor connectivity KPIs, confirming a mechanistic interpretation. Agradient-boosted classifier exploiting cross-KPI features recovers41–44% of the escalated band at 97% precision, raising totalautomation from 70% toward 80%. Feature ablation shows thatcross-KPI feature engineering, not classifier choice, drives thisgain.