---
title: "Prescriptive process monitoring: Quo vadis?"
type: source
tags: [prescriptive-process-monitoring, ppm, slr, framework, survey]
authors: [Kubrak, Kateryna; Milani, Fredrik; Nolte, Alexander; Dumas, Marlon]
year: 2022
venue: "PeerJ Computer Science 8:e1097"
kind: paper
raw_path: "raw/Predictive process monitoring/Prescriptive process monitoring - que vadiz.pdf"
doi: "10.7717/peerj-cs.1097"
created: 2026-04-20
updated: 2026-04-22
key_claims:
  - Prescriptive process monitoring (PrPM) is a family of methods that recommends interventions during the execution of a case so as to optimise the process with respect to an objective; it extends predictive process monitoring from forecasts to actions.
  - A systematic literature review following Kitchenham & Charters (2007) identified 37 relevant papers (1999–2021), characterised along a classification framework whose main axis is performance *objective* and whose six working dimensions are - performance objective, performance metric (target), intervention (type & process perspective), modeling technique, data input (perspective & feature encoding), and intervention policy (with sub-axes frequency and purpose).
  - The field splits cleanly into two objective camps — **optimising process outcome** (16 of 37 methods; binary/categorical case-level labels such as deadline violation or customer rejection) and **optimising process efficiency** (21 of 37; quantitative KPIs dominated by temporal metrics — cycle time and processing time are the two most over-represented targets).
  - Interventions concentrate on two process perspectives only - **control-flow** (next task / next sequence to perform) and **resource** (resource allocation / reassignment); other perspectives (data, time, compliance, cost redesign) are essentially absent, and interventions themselves are hand-specified by analysts rather than discovered from the log.
  - Modeling techniques split along the optimising/guiding axis - optimising methods use decision trees, random forests, LSTMs, SVMs, XGBoost, reinforcement learning, and — in a small but emerging sub-branch — causal inference (uplift/CATE models by Bozorgi et al. 2021, Shoush & Dumas 2021, Fahrenkrog-Petersen et al. 2022); guiding methods overwhelmingly use nearest-neighbour / similarity over historical traces.
  - Intervention policies take five concrete forms - similarity-based, rule-based, exceeding-metric-limits, maximum-metric-improvement, and probability-of-negative-outcome-above-threshold; orthogonally, policies differ along **frequency** (continuous vs discrete) and **purpose** (optimizing vs guiding, with optimizing further split into correlation- vs causality-based).
  - Six explicit research gaps ("Quo vadis?") - (i) lack of in-vivo validation (all 37 methods evaluated on historical or synthetic logs, only Dees et al. 2019 attempted real-world deployment and it did not produce desired outcomes), (ii) no principled method for *discovering* candidate interventions from event logs, (iii) policy design blind to causality and second-order effects, (iv) absence of explainability and human-in-the-loop feedback for both the underlying prediction and the triggering policy, (v) narrow performance coverage — temporal metrics dominate while quality, defect rate, revenue, compliance are under-served, and (vi) fragmented terminology ("proactive adaptation", "next-step recommendation", "next best action", "on-the-fly resource allocation" coexist without a common vocabulary).
  - Notably absent from the paper - any explicit reference to Hammer, Business Process Reengineering, Reijers & Liman Mansar's 29-heuristic redesign catalogue, the devil's quadrangle, BPM-lifecycle framing, change management, organisational-culture adoption barriers, or contingency-theoretic context-awareness. The only acknowledgement of implementation difficulty is the mention of "second-order effects" and the aside that detecting them "requires human judgment and iterative policy validation (e.g., via A/B testing)". PrPM is framed as a purely technical sub-field of process mining, detached from its design-time BPR predecessor.
---

# 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 [[sources/2007-kitchenham-slr-guidelines|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:

1. **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).
2. **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.
3. **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.
4. **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.
5. **Data input (perspective + feature encoding)** — four input perspectives: Control-flow, Resource, Temporal, Domain-specific. Feature encoding refines these (e.g., resource → experience / performance / workload).
6. **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:

1. **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.
2. **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.
3. **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.
4. **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.
5. **Breadth of performance objectives** — temporal metrics dominate; quality, defect rate, revenue, compliance are under-represented. The paper argues for broadening the objective space.
6. **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 [[sources/2005-reijers-limanmansar-best-practices-bpr|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 [[sources/2021-dumas-process-mining-2-from-insights-to-action|Dumas 2021]] and picks up the programme initiated a decade earlier by [[sources/2014-groger-prescriptive-analytics-bpo|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 [[syntheses/bpm-phases-and-bpr-legacy#interlude--does-prpm-repeat-or-learn-from-bpr|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]]