LCDD: Detecting Business Process Drifts Based on Local Completeness

A Survey on Concept Drift in Process Mining

“…The authors in [40] performed several experiments to evaluate the accuracy of LCDD using the ML equation to support the definition of the CW parameter. The accuracy of the fixed windows obtained using synthetic logs is considerably higher than other approaches: Runs and Alpha options from Apromore ProDrift [45,57], and TPCDD [88].…”

“…Yet, tuning this parameter can result in higher accuracy when comparing to the default parameters applied to the compared approaches. Authors in [40] also report a real-world experiment comparing LCDD performance against the same methods applied to the synthetic datasets. Because LCDD proved to be not robust to noise, the authors apply the method defined in [21] to obtain similar results to the compared methods.…”

“…We suggest that the TPs can only score CPs detected after the real CP inside a neighborhood pre-defined. The idea of an Error Tolerance (ET), described in [88], similar to the objective evaluation using a lag period proposed in [40,51], can be adapted to define this neighborhood. The main difference between the current proposals is that we argue that the lag period should be considered only after the real drift position and not before it, as drift detection is a reactive process rather than a preemptive one.…”

“…For instance, if the method returns the trace index as the CP, a TP can consider a range of indexes around the actual drift, and this should be specified. In [40,51], the authors clearly define the evaluation mechanism (Figure 7). The difference from other papers is that the authors apply a lag period l surrounding a detected or an actual drift.…”

“…The approach proposed in [40] applies the local completeness (LC) property of event logs defined in [81] for drift detection. The idea is that it is possible to assert (with a defined confidence level K) that a log satisfies LC in a limited length.…”