---
title: "Prescriptive Analytics for Recommendation-Based Business Process Optimization"
type: source
tags: [prescriptive-process-monitoring, process-warehouse, data-mining, recommendation, manufacturing]
authors: [Gröger, Christoph; Schwarz, Holger; Mitschang, Bernhard]
year: 2014
venue: "BIS 2014, LNBIP 176 (Springer), pp. 25–37"
kind: paper
raw_path: "raw/Predictive process monitoring/prescriptive analytics for recommendation-based business process optimization (2014).pdf"
created: 2026-04-20
updated: 2026-04-20
key_claims:
  - Existing business activity monitoring stops at descriptive / predictive analytics and leaves the step of translating analysis results into improvement actions to subjective user judgment; the proposed recommendation-based Business Process Optimization (rBPO) fills this gap with prescriptive analytics.
  - rBPO sits on top of a holistic process warehouse (PWH) that integrates process data (flow-oriented - events, models) with operational data (subject-oriented - resources, machines, ERP attributes) and stores mining models for reuse.
  - The runtime procedure has two steps - real-time prediction classifies the running instance's likely metric outcome (OK / NotOK) with a decision tree after each process step; if NotOK, recommendation generation induces a second decision tree over only those completed instances whose preceding attributes match, and reads off "OK"-labelled paths as decision rules to recommend action settings for the next step.
  - Restricting the training set to past instances that share the running case's completed-step attribute values is essential to respect implicit process restrictions (e.g., material constraints on machines); using all historical instances yields recommendations that violate such restrictions.
  - The approach is demonstrated in manufacturing (steel-spring production) but presented as a universal concept applicable wherever a suitable process warehouse exists — proof of concept via a prototypical implementation.
---

# 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:

1. **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.
2. **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 [[sources/2022-kubrak-prescriptive-ppm-slr|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 — [[sources/2021-dumas-process-mining-2-from-insights-to-action|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).