Kubrak, Milani, Nolte & Dumas 2022 — Prescriptive Process Monitoring: Quo Vadis?

Systematic literature review (SLR) of the prescriptive process monitoring (PrPM) field. Published in PeerJ Computer Science (September 2022). The definitive map of the PrPM sub-field up to mid-2021, and the scaffolding for the concept page concepts/prescriptive-process-monitoring.

Method

The authors follow Kitchenham & Charters (2007): they formulated five research questions (RQ1 objective, RQ2 interventions, RQ3 data, RQ4 modeling techniques, RQ5 policies), searched ACM DL, Scopus (incl. SpringerLink), Web of Science and IEEE Xplore with two complementary search strings — (recommender OR "decision support" OR prescriptive) AND "process mining" and (recommender OR "next activity" OR "next step" OR "next resource" OR proactive) AND "business process" — applied four exclusion criteria (digitally accessible, English, non-duplicate, ≥6 pages) and three inclusion criteria (relevance to PrPM, describes a method or a case, identifies at least one way to elicit candidate interventions for an ongoing case), and snowballed backward. 1,367 candidate papers funnelled to a final corpus of 37 papers (earliest 2008; 20 of 37 published 2017–2021). Eight journal articles, 29 conference papers. 19 real-life logs, 9 synthetic, 3 both.

The classification framework — six working dimensions

Although the full framework lists ten columns (including detail-of-algorithm and running example), the conceptual scaffolding collapses to six dimensions that classify the 37 methods:

Performance objective — what the method optimises. Two camps:
- Process outcome (16 methods) — probability of a desired case-level label, typically binary (deadline violated / not, customer rejects delivery / not, critical patient state / not).
- Process efficiency (21 methods) — a quantitative KPI at the case level (cycle time, processing time, labour cost, revenue, defect rate).
Performance metric / target — the concrete KPI operationalising the objective. Most skewed dimension: temporal metrics (cycle time + processing time + deadline violation) cover roughly two-thirds of all methods; quality, revenue and compliance targets are represented by single-digit counts.
Intervention type (process perspective + concrete action) — mostly control-flow interventions (recommend next task / next sequence of tasks) or resource-allocation interventions (assign a resource or reassign a task). Few methods prescribe interventions on data, time, or other perspectives; all interventions are hand-specified by analysts — no method discovers them from the log.
Modeling technique — decision trees / random forests, nearest-neighbour / similarity, LSTMs / RNNs, SVMs, XGBoost, reinforcement learning (notably Metzger, Kley & Palm 2020), and causal inference (CATE / uplift models in Bozorgi et al. 2021, Shoush & Dumas 2021, Fahrenkrog-Petersen et al. 2022). Detail-of-algorithm is sufficient for reproduction in 76% of papers.
Data input (perspective + feature encoding) — four input perspectives: Control-flow, Resource, Temporal, Domain-specific. Feature encoding refines these (e.g., resource → experience / performance / workload).
Intervention policy — the trigger rule. Five concrete forms:
- similarity-based (guiding methods — nearest-neighbour over past traces),
- rule-based / set-of-rules,
- exceeding-metric-limits,
- maximum-metric-improvement,
- probability-of-negative-outcome-above-threshold. Orthogonally, policies are characterised along two sub-axes:
- frequency: continuous (prescribe at every step) vs discrete (only when a trigger fires),
- purpose: optimizing (act to improve a KPI, correlation- or causality-based) vs guiding (imitate successful past cases via similarity).

Dominant approaches — quantitative picture

Guiding-by-similarity is the modal method. Nearest-neighbour / trace-similarity is the single most-used technique, dominating the "guiding" sub-family (Mertens 2020; Haisjackl & Weber 2010; Arias et al. 2018; Arias, Muñoz-Gama & Sepúlveda 2016; Nezhad & Bartolini 2011; Triki et al. 2013; Yang et al. 2017; Thomas, Kumar & Annappa 2017; Terragni & Hassani 2019).
Correlation-based optimising methods cluster around decision trees and classical ML classifiers (Gröger, Schwarz & Mitschang 2014; Sindhgatta, Ghose & Dam 2016; Conforti et al. 2015; Ghattas, Soffer & Peleg 2014; Kim, Obregón & Jung 2013; Obregón, Kim & Jung 2013; de Leoni, Dees & Reulink 2020).
Neural-sequence optimising methods are a 2019–2021 wave (Park & Song 2019; Weinzierl et al. 2020a/b; Khan et al. 2021; Metzger, Kley & Palm 2020 — the latter RL + LSTM hybrid).
Causality-based optimising methods form a small but argued-as-most-promising sub-branch (Shoush & Dumas 2021; Bozorgi et al. 2021; Fahrenkrog-Petersen et al. 2022; Teinemaa et al. 2018). The paper explicitly flags that "only a few existing methods take causality into account when designing policies" and calls this out as a research direction.
Optimisation / constraint programming appears in a handful (Schonenberg et al. 2008; Barba, Weber & Valle 2011).
Reinforcement learning appears in one method (Metzger, Kley & Palm 2020).

"Quo vadis?" — six open problems

Condensed from §Research Gaps and Implications:

In-vivo validation — only one paper in the corpus (Dees et al. 2019) attempted real-world deployment; predictions were accurate but interventions did not produce desired outcomes. Everything else is back-tested on historical logs.
Discovery of candidate interventions — no method systematically elicits interventions from event logs, textual documentation, or structured process metadata. Discrete methods leave intervention design to stakeholders a priori; continuous methods restrict themselves to next-task or next-resource recommendation.
Causality-aware policy design — correlation-based prescriptions do not address the cause of a negative outcome. The paper argues for causal-estimation modules (CATE / uplift) as a first-class policy-design primitive.
Explainability and feedback loops — no method in the corpus explains why an intervention is recommended, either at the prediction layer or at the policy layer. Human-in-the-loop integration is absent.
Breadth of performance objectives — temporal metrics dominate; quality, defect rate, revenue, compliance are under-represented. The paper argues for broadening the objective space.
Common terminology — "proactive process adaptation", "on-the-fly resource allocation", "next-step recommendation", "next best action" coexist without a shared vocabulary.

The authors also note, in passing, that second-order effects (an intervention causing a downstream consequence) require human judgment and A/B testing — effectively gesturing at empirical-implementation science without naming it.

What the paper does not discuss

A close reading reveals that Kubrak et al. treat PrPM as a self-contained branch of process mining with no historical root-system. Specifically, the text contains zero mentions of:

Hammer, Business Process Reengineering, BPR, reengineering (the sole "BPR" string is the algorithm name Bayesian Personalized Ranking in a framework cell),
Reijers & Liman Mansar's 29 redesign heuristics (the only two "Reijers" hits are author citations of unrelated works — Dees et al. 2019 and Dumas et al. 2018),
the devil's quadrangle (time / cost / quality / flexibility trade-off),
the BPM lifecycle (Identify → Discover → Analyse → Redesign → Implement → Monitor),
change management, adoption barriers, organisational culture, or contingency-theoretic context-awareness,
management-science traditions (TQM, Lean, Six Sigma) or decision-support-systems lineage.

This absence is itself a finding. The one-and-only implementation-side acknowledgement is the observation that detecting second-order effects "requires human judgment and iterative policy validation (e.g., via A/B testing)". The PrPM field, as Kubrak portrays it, has been engineered without a design-time predecessor — a striking omission given that Reijers & Liman Mansar (2005) catalogued exactly the same class of interventions (resource allocation, task reordering, control-flow redesign) two decades earlier, and explicitly evaluated each against the devil's quadrangle.

Why this page matters

This is the anchoring SLR for the prescriptive end of the descriptive → predictive → prescriptive analytics pyramid in BPM. It supersedes and formalises the vision sketched by Dumas 2021 and picks up the programme initiated a decade earlier by Gröger, Schwarz & Mitschang 2014. Its six-dimensional framework is the scaffolding for the concept page concepts/prescriptive-process-monitoring.

Its silence on BPR is analysed in bpm-phases-and-bpr-legacy §Interlude.

Connections

Concepts: concepts/prescriptive-process-monitoring · concepts/predictive-process-monitoring · concepts/intervention-policy · concepts/process-mining-spectrum · concepts/operational-support · concepts/causal-process-discovery · concepts/bpr-heuristics · concepts/devils-quadrangle · concepts/context-aware-bpm · concepts/bpm-lifecycle

Methods: methods/systematic-literature-review

Entities: entities/kateryna-kubrak · entities/fredrik-milani · entities/alexander-nolte · entities/marlon-dumas · entities/university-of-tartu

Related sources: sources/2014-groger-prescriptive-analytics-bpo (early prescriptive concept — cited by Kubrak as correlation-based optimising method) · sources/2021-dumas-process-mining-2-from-insights-to-action (keynote vision) · sources/2007-kitchenham-slr-guidelines (SLR method) · sources/2005-reijers-limanmansar-best-practices-bpr (design-time predecessor — not cited by Kubrak; the omission is itself analytically significant) · sources/2018-dumas-fundamentals-of-bpm (cited by Kubrak for the quality / efficiency split).

Syntheses: syntheses/bpm-phases-and-bpr-legacy · syntheses/ppm-landscape