Introduction
Traditional process mining often operates under a binary assumption: either the event log (the recorded reality) or the process model (the intended design) is a perfect ground truth. In practice, however, both sources are frequently flawed. Event logs may contain noise or sensor errors, while models may be outdated or poorly designed. This paper introduces a generalized conformance checking framework that acknowledges the inherent uncertainty in both domains, allowing users to reconcile and repair logs and models simultaneously through adjustable "trust levels."
Methodology: A Two-Step Reconciliation
To manage the computational complexity of simultaneously modifying both the log and the model, the authors propose a "divide-and-conquer" approach. This heuristic search strategy operates across two distinct spaces:
- Model-Space Approximation: The framework first "lifts" the event log into the model space by mining a representative model, $M(L)$. A greedy search then identifies an intermediate repaired model, $M^*$, positioned between the original input model and the mined model. This position is governed by a user-defined Model Trust level ($\pi_M$).
- Log-Space Alignment: With $M^*$ established, the original event log is aligned to this new model. A Log Trust level ($\pi_L$) determines the extent to which the log is repaired to match the model versus being flagged as a persistent misalignment.
The Trust Spectrum
A key contribution of this work is the conceptualization of process mining tasks as points on a continuous spectrum of trust. By treating trust as a parameter, the authors unify several previously distinct tasks into a single mathematical framework.
As shown in the mapping below, traditional tasks represent specific boundary conditions of this generalized model:
- Classical Process Discovery: Absolute trust in the log ($\pi_L=1$) and zero trust in the model ($\pi_M=0$).
- Conformance Checking: Absolute trust in both sources ($\pi_L=1, \pi_M=1$).
- Garbage In, Garbage Out: Zero trust in either source ($\pi_L=0, \pi_M=0$).
Key Results and Insights
The framework successfully categorizes misalignments into three buckets: log anomalies, modeling errors, and unresolvable inconsistencies. By sequentializing the search—identifying model changes before log repairs—the authors provide a computationally tractable solution to a complex problem. While specific benchmark data is detailed in the full paper [Ro16], the authors report that this generalized approach achieves model quality comparable to state-of-the-art repair techniques while offering significantly greater flexibility.
Limitations and Future Work
The current approach relies on a heuristic "divide-and-conquer" method, which may not always reach a global mathematical optimum compared to a simultaneous search. Furthermore, the framework assumes that users can accurately provide trust levels ($\pi_L$ and $\pi_M$). Future research aims to automate the estimation of these trust levels using objective data quality metrics and to refine the distance measures used for replay fitness.
Summary updated