Intervention Policy

In prescriptive process monitoring (PrPM), an intervention policy specifies when an intervention is triggered, to which running cases it is applied, and how the candidate intervention is chosen. The policy is orthogonal to the prediction model and the catalogue of candidate interventions — it is the decision rule sitting on top of both.

Kubrak, Milani, Nolte & Dumas (2022) elevate intervention policy to one of six first-class dimensions of their PrPM framework (alongside performance objective, performance metrics, intervention types, modeling techniques, and data inputs).

Axes of variation

Frequency — how often does the policy fire?

Continuous — an intervention is prescribed for multiple (possibly every) step of an ongoing case. Typical in "guiding" recommender systems.
Discrete — an intervention is triggered only when a specific condition is detected (e.g., predicted negative outcome above a threshold). Typical in alarm-based systems.

Purpose — what is the policy's logic?

Optimizing — trigger an intervention so as to improve a KPI. Sub-split:
- Correlation-based — pick the intervention that historically co-occurs with the best outcomes (e.g., Gröger et al. 2014).
- Causality-based — pick the intervention that is causally estimated to improve the outcome (uplift trees, CATE estimation; Bozorgi et al. 2021; Shoush & Dumas 2021).
Guiding — prescribe recommendations based on similarity to historical traces, without an explicit KPI target (nearest-neighbour recommenders).

Common trigger types

Threshold on predicted probability of a negative outcome (alarm-based; Teinemaa, Fahrenkrog-Petersen et al.).
Set of rules over process / resource attributes.
Maximum expected metric improvement — assign intervention to the case/resource pair that maximises predicted KPI delta.
Cost-aware alarm — fire only when expected benefit exceeds intervention cost (Fahrenkrog-Petersen et al. 2022's Fire now, fire later).
Similarity to past successful traces — pure guiding.

Why the distinction matters

A method with an excellent predictor but a poorly designed policy will waste interventions (firing when it shouldn't) or miss opportunities (not firing when it should). The predictor answers what will happen; the policy answers what to do given the prediction, the cost of acting, and the limits on resources.

Causality and second-order effects

The Kubrak SLR names causal, second-order-aware policy design as the most promising research direction. Correlation-based policies can recommend interventions that would occur anyway in the successful cases (selection bias) or that conflict with implicit process restrictions (as Gröger et al. observed with material-machine compatibility). Causal methods (uplift modelling, counterfactual reasoning; cf. concepts/causal-process-discovery) aim to estimate the true effect of intervening, and second-order analysis considers how an intervention on one case affects contemporaneous cases through shared resources.

Resource constraints

Practical policies must account for resource availability — Shoush & Dumas (2021) combine probability-of-undesired-outcome with resource availability and intervention cost before prescribing. Without resource-awareness, a policy can fire more interventions than can be executed.

concepts/prescriptive-process-monitoring · concepts/predictive-process-monitoring · concepts/causal-process-discovery · concepts/interventionist-theory-of-causation (philosophical grounding: Woodward's causes-as-handles view is the metaphysical counterpart of PrPM policy design) · concepts/operational-support · concepts/mape-k-loop (the Plan stage of MAPE-K is the APM analogue of a PrPM policy)