A Recovery Module for an Incident Management Assistant with Missing and Imbalanced Data

• Alisson N. Bonato Amadeus IT Group
• Jade S. Hatanaka Marques Amadeus IT Group / UFRGS
• Rajnish Kumar Amadeus IT Group

DOI: https://doi.org/10.5753/eramiars.2025.16747

Resumo

In Information Technology (IT) Service Management, modern approaches leveraging Artificial Intelligence for IT Operations (AIOps) typically rely on abundant, high-quality observability data. However, many real-world environments operate under severe data constraints, such as scarce historical records and incomplete information. This paper introduces AIMA+, an incident management assistant designed specifically for these challenging conditions. Given an incident’s textual description as input, AIMA+ outputs a ranked list of predicted recovery actions, retrieves similar past incidents, and generates a synthesized textual summary to guide operators. Our core contribution is a methodology that addresses data incompleteness through a Large Language Model (LLM)-based data augmentation strategy and data scarcity with an interpretable, multi-pronged framework. Our experiments show that augmenting the dataset with expert-guided LLM classifications dramatically improved predictive performance, increasing the macro F1-score from a baseline of 0.2 to 0.6. This work presents a pragmatic blueprint for developing effective AIOps decision support tools in realistic, imperfect industrial settings.

Palavras-chave: Incident Management, AIOps, Recovery Actions, Imbalanced Data, Large Language Models, Multilabel Classification

Referências

Chen, Z., et al. (2023). Chat-based Incident Triage with Retrieval Augmented Generation. In ESEC/FSE.

Du, M., et al. (2017). DeepLog: Anomaly Detection and Diagnosis from System Logs through Deep Learning. In CCS.

Nguyen, T. T., et al. (2012). Recommending Similar Bug Reports and Their Fixes. In ICSE.

Wang, X., et al. (2019). Learning to Predict and Mitigate Push-Caused Production Failures. In OSDI.

Yuan, Z., et al. (2021). NetSieve: A Scalable and Robust Causal Inference Framework for Incident Analysis. In SOSP.

Zhang, X., et al. (2019). LogRobust: A Robust Online System Log Anomaly Detection System. In USENIX Security Symposium.