Case Management: Lifecycle Data for Prediction
Case Management
Case Management tests how work is modeled and how case data becomes analytical features and labels. Expect case types, stages, steps, processes, child cases, data propagation, service levels, and case lifecycle. For a data scientist, case design determines what is captured, when it is captured, and whether an outcome is observable.
Core rules
- A case type defines the lifecycle: stages contain steps, and steps invoke processes or assignments.
- Child cases encapsulate sub-work; data does not automatically propagate to the parent unless configured.
- Service levels measure goals and deadlines; they create timestamps useful for SLA prediction.
- Case status, stage, and step are operational states, not necessarily clean analytical labels.
- Define the label at a point in time and avoid features created after that point.
Worked example
Predict whether a case will miss its SLA. Reasoning: use only data available at case creation or at a defined checkpoint. Features: channel, priority, case type, and initial document count. Label: SLA breached at resolution. Exclude reassignment count after the checkpoint because it is post-outcome. Validate that resolved and open cases are handled correctly; open cases are censored and should not be labeled as not breached.
Traps
- Using final case status as a feature instead of a label.
- Assuming child case properties are visible in the parent without propagation.
- Mixing stage names with status values; they are different.
- Ignoring censoring: open cases have unknown outcomes.
- Building a model on data captured only after a manual step that may not occur.