Predictive Framework for Decision-Making Under Uncertainty in the Handling of Failures in IT Infrastructures

This paper proposes an approach for anticipating critical-state entry in IT infrastructures from log-derived time series. The representation combines failure rate, critical episodes, operational states, and causal temporal features in a reproducible pipeline aimed at building a supervised anticipation dataset. In the primary TimeSeriesSplit validation, Logistic Regression achieved a mean recall of 0.784, a mean F1-score of 0.478, and a mean ROC-AUC of 0.640, indicating the capture of precursor signals despite relevant temporal instability. The findings suggest that problem formulation plays a central role in predictive performance and that model outputs should be interpreted as relative temporal risk scores for decision support under uncertainty.