Improving Heuristic Process Discovery Methods Through Determining the Optimal Split/Join Patterns of Dependency Graphs

Abstract

Identifying the split and join patterns of dependency graphs is an essential step in Heuristics Mining process discovery methods. The existing methods determine the split/join patterns (consisting of AND and XOR relations) according to the event log information about the activities involved in the splits and joins. Hence, they neglect the event log information available for the other activities on the paths from split points to join points. On the other hand, the current methods determine the patterns of each split/join separately and do not aim to select the best set of patterns. Therefore, the outputs of the existing methods can be non-optimal. Furthermore, the current methods cannot guarantee that there is a matching And-join for each AND-split, and vice versa. This can make some split/join patterns incapable of being activated. To handle these issues, this paper, for the first time, presents an integer linear programming model which identifies the optimal patterns of splits/joins with regard to all succession information that is available in the event log; simultaneously, it ensures that for each AND-split there is at least one matching AND-join, and vice versa. The objective function of the proposed model is inspired by replay fitness and precision dimensions of process model quality. According to the assessments, the process models obtained by the proposed method are superior to the results of the most prominent methods of determining split/join patterns in terms of replay fitness, precision, simplicity, and matching AND-splits with AND-joins.

Keywords

• Heuristics Miner
• integer linear programming
• process discovery
• process mining
• mining split/join patterns
• optimization