Explainable Hybrid Feature Selection and Stacked Ensemble Learning for Intelligent Incident Prioritization in IT Service Management

Abstract

Effective incident management is a key element in IT Service Management (ITSM), where the level of service incident prioritization accuracy has an immediate impact on system availability, response time, and fulfillment of Service Level Agreements (SLAs). Traditional rule-based prioritization methods usually cannot reflect the complex interaction between attributes of incidents, causing the workload of prioritization to be high and the allocation of resources to be unbalanced. To overcome this drawback, an interpretable hybrid machine learning model is put forward for automated prediction of incident priority in the context of ITSM. The proposed method combines hybrid feature selection with Extra Trees followed by Mutual Information and SMOTE is used to handle class imbalance. A stacked ensemble model consisting of LightGBM, CatBoost and Random Forest model is trained as a meta-learner which is Logistic Regression to output the final predictions. To improve transparency and interpretability, SHAP (SHapley Additive exPlanations) is added to explain the MGM model globally and locally, thus explaining the decision-making process of the model. The proposed approach was tested on the public ITSM incident data, which consists real world service desk records and has various incident attributes. Results show that the hybrid ensemble framework outperforms the single based machine learning models substantially in terms of prediction performance. The method proposed obtained a total classification accuracy of 0. 9998 with the precision, recall and F1-scores being 1.00 in the majority of the priority groups.  The confusion matrix showed almost perfect classification, detecting most of the incident priority levels with very few misclassifications. Particularly, the model is able to predict Priority 1 (139 samples), Priority 2 (1065 samples), Priority 3 (4544 samples), Priority 4 (3297 samples) and Priority 5 (276 samples) with precision and recall of 1.00 for both suggesting a very good robustness when considering both majority and minority class. Cross-validation results also verified the robustness of the proposed hybrid model. The add-in of some explainable AI approaches yields de-facto actionable explanation on what drives incident prioritization, allowing IT manager so gain insights on the effect of such factors as urgency, impact, and complexity of the incident. These results endorse the effectiveness of the proposed hybrid ensemble framework for enabling intelligent decision making in IT service operation through providing precise, transparent, and automated incident prioritization. This method offers a promise to greatly enhance service desk productivity, decrease response time, and improve business processes in today’s IT infrastructure environment.