Temporal Heterogeneous Graph Neural Networks For Incident Assignee Recommendation

Our work presents an approach for recommending incident assignees in large-scale IT service management (ITSM) systems using a heterogeneous graph neural network (GNN). Incident assignment is a key step in operational workflows, as the choice of assignee directly affects resolution time, workload distribution, and service quality. We compare different temporal modeling approaches, such as daily snapshots, in which each day’s incident data is represented as a separate heterogeneous graph and time-stamped edge attributes. The graph reflects the relationships already present in the ITSM records. This gives us a structured picture of how the different pieces of information relate to each other. We assume that temporal information plays an essential role in identifying incident assignment patterns, such as changing workloads, shifting team structures, and asset dependencies. By incorporating temporal dynamics through sequential daily snapshots, the model maintains and updates node representations over time. The results show that the temporally aware heterogeneous GNN captures how relationships in the ITSM data change over time and that it boosts the accuracy of incident-to-assignee recommendations noticeably compared to models that do not account for temporal dynamics.