Paper 03 — Conformance Checking under Concept Drift: A Stream-Based Approach for Long-Running Business Processes

Paper numberCONFENIS 2027 / 03

TrackProcess Mining and AI

Pages31–45

ProceedingsSpringer LNBIP vol. 528 (2027)

Correspondingj.koehler@ist.tugraz.at

DOI10.1007/978-3-031-XXXXX-X_3

Abstract

Real-world business processes evolve continuously: regulatory updates, market shifts and organisational learning all induce gradual or sudden changes to the underlying process model. Classical conformance checking, formulated as a batch problem against a fixed reference model, is poorly equipped to handle this drift. We propose CC-Drift, a stream-based conformance-checking framework that maintains an online alignment-cost estimator and explicitly models drift via a sliding ADWIN window. We evaluate CC-Drift on three real-world event streams totalling 14 million events and find that it detects sudden drift within an average of 1,820 events of onset, compared with 14,400 events for a periodic batch baseline.

Keywords: Conformance Checking; Concept Drift; Streaming Process Mining; Online Learning; Long-Running Processes

1. Introduction

Conformance checking — quantifying the deviation between an observed event log and a reference process model — is one of the three pillars of process mining (van der Aalst, 2016). The dominant alignment-based formulation (Adriansyah et al., 2011) assumes a stationary process: the reference model is fixed, and the log is processed in batch. This assumption is increasingly untenable. Long-running processes such as procure-to-pay in multinational enterprises, claims handling in insurance, or patient pathways in healthcare evolve under regulatory and organisational pressure (Bose et al., 2014; Maaradji et al., 2017). Concept drift — change in the data-generating process — produces false-positive conformance violations against an out-of-date reference and obscures genuine compliance regressions against an emerging norm.

2. Method

CC-Drift maintains three streaming structures: (1) an online alignment-cost estimator based on the prefix-alignment approach of van Zelst et al. (2019), updated per event with amortised O(|model|) cost; (2) an ADWIN (Adaptive Windowing) change detector (Bifet & Gavaldà, 2007) applied to the alignment-cost stream; and (3) a model-rediscovery trigger that fires an inductive-miner re-run on the post-drift window when ADWIN signals a change point with confidence δ = 0.001. The framework is implemented in Python 3.11 on top of PM4Py 2.7 and the river streaming library, with Kafka as the input transport. Latency on a single 16-core machine averaged 0.42 ms per event for models with up to 120 activities.

3. Results

Evaluation used three event streams: (S1) a public BPI Challenge 2017 loan-application log replayed at 100× wall-clock speed with two synthetic drift injections; (S2) a production order-to-cash stream from a partnering automotive supplier (4.1 M events, real drift in May 2025 following an SAP enhancement-pack upgrade); and (S3) a healthcare patient-pathway stream from Steiermärkische Krankenanstaltengesellschaft (9.8 M events, real drift in January 2026 following ICD-11 adoption). CC-Drift detected sudden drift on S1 at events 1,710 and 1,930 after injection, against a ground-truth detection budget of 5,000 events (improving on the 14,400-event mean of the daily-batch baseline). On S2 and S3 we have no ground-truth onset, but CC-Drift signalled change points within 24 hours of the operationally known events, against batch baselines of 28 and 35 days. False-positive rates were 0.7% over 14 days of post-drift stable operation, comparable to ADWIN's published behaviour.

4. Discussion

CC-Drift offers a substantive reduction in time-to-detection for drift in long-running processes, at modest computational cost. Two limitations bound its applicability. First, ADWIN's confidence parameter must be tuned for each domain; too-aggressive settings produce spurious model rediscovery and operational churn. Second, the prefix-alignment cost estimator is an approximation; on heavily concurrent models, our estimator deviates from the optimal alignment by up to 8% in synthetic stress tests.

5. Conclusion

Stream-based conformance checking is feasible and operationally valuable for long-running EIS processes subject to drift. CC-Drift offers a baseline framework for practitioners and a benchmark for further research on online process mining.

References

Adriansyah, A., van Dongen, B.F., van der Aalst, W.M.P.: Conformance checking using cost-based fitness analysis. In: IEEE EDOC 2011, pp. 55–64 (2011).
Bifet, A., Gavaldà, R.: Learning from time-changing data with adaptive windowing. In: SDM 2007, pp. 443–448 (2007).
Bose, R.P.J.C., van der Aalst, W.M.P., Žliobaitė, I., Pechenizkiy, M.: Dealing with concept drifts in process mining. IEEE TNNLS 25(1), 154–171 (2014).
Maaradji, A., Dumas, M., La Rosa, M., Ostovar, A.: Detecting sudden and gradual drifts in business processes from execution traces. IEEE TKDE 29(10), 2140–2154 (2017).
van der Aalst, W.M.P.: Process Mining: Data Science in Action, 2nd edn. Springer (2016).
van Zelst, S.J., van Dongen, B.F., van der Aalst, W.M.P.: Online conformance checking: relating event streams to process models using prefix-alignments. International Journal of Data Science and Analytics 8(3), 269–284 (2019).

Citation: Johannes Köhler, Anna Müller. "Conformance Checking under Concept Drift: A Stream-Based Approach for Long-Running Business Processes." In: Tjoa, A.M., Mendling, J., Wimmer, M. (eds.) Research and Practical Issues of Enterprise Information Systems. CONFENIS 2027. LNBIP 528, pp. 31–45. Springer, Cham (2027).

← Back to proceedings