Gröger, Schwarz & Mitschang 2014 — Prescriptive Analytics for Recommendation-Based Business Process Optimization

Early paper coining the prescriptive-analytics angle for BPM, published in the BIS 2014 proceedings (LNBIP 176). Authors from the Institute of Parallel and Distributed Systems, University of Stuttgart.

Thesis

Standard business-activity monitoring answers what happened (descriptive) and, with PPM, what is likely to happen (predictive). It does not answer what should be done. The paper proposes recommendation-based Business Process Optimization (rBPO) — a runtime loop that, upon predicting an impending metric deviation (e.g., cycle time exceeding a threshold), emits a concrete action recommendation (e.g., "use machine M33 and perform a reset before peening"). The authors place this explicitly on the prescriptive-analytics side of the descriptive → predictive → prescriptive pyramid.

Architecture (§2, Fig. 1)

rBPO sits on top of a holistic Process Warehouse (PWH) that integrates:

Process data — flow-oriented: events, process-model data.
Operational data — subject-oriented: employee, machine, tool attributes, settings.

Each completed process instance is stored denormalised as one row with attributes for every step plus a categorised metric label (OK / NotOK) defined by an analyst. Per-step, a two-phase loop runs:

Real-time prediction (§3) — train a binary decision tree over completed-step attributes of past instances; apply to the running instance to classify its predicted end-metric.
Recommendation generation (§4) — if NotOK, induce a second decision tree whose data basis is restricted to past instances with identical completed-step attribute values, using only directly actionable attributes of the next step (e.g., machine ID, tool ID, settings) as independents. Paths ending in "OK" leaves become decision rules; these are read out as an action recommendation.

Decision trees are chosen over association rules for comprehensibility and over n-ary trees are preferred binary for simplicity. Recommendations are explicitly multi-action (R5).

Why restricted training matters (§4.1)

Using all past completions gives recommendations that ignore implicit process restrictions (e.g., certain materials are only compatible with certain machines). Restricting to instances with matching completed-step attributes lets the decision tree discover such dependencies. The paper illustrates the failure mode with a toy example where the unrestricted variant recommends a machine incompatible with the running case's material.

Proof of concept (§5)

Demonstrated on a steel-spring manufacturing process (winding → tempering → shot peening). Authors claim domain-independence — rBPO works for any domain with a suitable PWH, including workflow-based processes.

Positioning in the PPM/PrPM lineage

This is the earliest clear articulation of prescriptive process monitoring in the BPM literature. Cited by Kubrak et al. 2022 as one of the correlation-based optimizing methods in the PrPM SLR. Predates the "alarm-based prescriptive monitoring" line (Teinemaa, Fahrenkrog-Petersen) and the causal-PrPM line (Bozorgi, Shoush) by 5–7 years. Conceptually aligned with — but independent of — Dumas 2021's "Process Mining 2.0" vision.

Connections

Concepts: concepts/prescriptive-process-monitoring · concepts/predictive-process-monitoring · concepts/operational-support · concepts/intervention-policy · concepts/trace-encoding

Entities: entities/christoph-groger · entities/university-of-stuttgart

Related sources: sources/2022-kubrak-prescriptive-ppm-slr (SLR that catalogues this paper) · sources/2021-dumas-process-mining-2-from-insights-to-action (vision statement from the Tartu school) · sources/2005-reijers-limanmansar-best-practices-bpr (design-time predecessor; same intent — recommend actionable process improvements — but at design rather than runtime).