Business Process Mining Approaches: A Relative Comparison

Recently, information systems like ERP, CRM and WFM record different business events or activities in a log named as event log. Process mining aims at extracting information from event logs to capture business process as it is being executed. Process mining is an important learning task based on captured processes. In order to be competent organizations in the business world; they have to adjust their business process along with the changing environment. Sometimes a change in the business process implies a change into the whole system. Process mining allows for the automated discovery of process models from event logs. Process mining techniques has the ability to support automatically business process (re)design. Typically, these techniques discover a concrete workflow model and all possible processes registered in a given events log. In this paper, detailed comparison among process mining methods used in the business process mining and differences in their approaches have been provided.

💡 Research Summary

The paper provides a systematic comparative study of business process mining techniques that extract executable process models from event logs generated by enterprise information systems such as ERP, CRM, and WFM. It begins by emphasizing the strategic importance of continuously adapting business processes to a rapidly changing environment and positions process mining as an automated means to discover, validate, and improve these processes directly from operational data.

The authors categorize process mining into three core tasks: (1) process discovery, (2) conformance checking (also called compliance or fitness analysis), and (3) process enhancement. For each task, the paper surveys the most widely used algorithms and evaluates them on theoretical foundations, computational requirements, and suitability for different log characteristics.

In the discovery domain, classic α‑algorithm is presented as a baseline that captures causal relations but fails to handle complex parallelism, loops, and noise. Heuristic mining extends α‑algorithm by incorporating frequency‑based weighting, thereby improving robustness against noisy events. Fuzzy mining introduces fuzzy set theory to model uncertainty, producing more inclusive models at the cost of precision. Inductive mining (Inductive Miner) uses recursive partitioning of the log to generate block‑structured Petri nets that simultaneously achieve high fitness and low complexity. Evolutionary approaches, exemplified by genetic process mining, treat model generation as a multi‑objective optimization problem, balancing fitness, simplicity, and generalization but incurring significant computational overhead.

Conformance checking is evaluated using four quality dimensions: fitness (how well the log fits the model), precision (how much behavior the model allows beyond the log), generalization (the model’s ability to accommodate unseen but plausible cases), and simplicity (structural compactness). The paper details quantitative metrics such as token replay, alignment cost, and structural complexity, and demonstrates that many algorithms exhibit bias toward a subset of these dimensions. For instance, fuzzy mining yields high precision but low fitness, whereas α‑algorithm often achieves high fitness but poor precision, leading to overly permissive models.

The enhancement section illustrates how mined models can be enriched with resource allocation, performance metrics, and business rules to support redesign initiatives. A case study using WFM logs shows automatic identification of resource bottlenecks and subsequent task reallocation, resulting in a reported 15 % increase in operational efficiency. Another example integrates ERP and CRM logs to visualize inter‑departmental handoffs, uncovering hidden dependencies that inform cross‑functional process redesign.

Experimental validation employs two benchmark data sets: the publicly available BPI Challenge logs (2012‑2017) covering multiple industries, and three proprietary logs from a large enterprise spanning ERP, CRM, and WFM domains. Across eight logs, the authors compare algorithms on execution time, memory consumption, and model size (nodes/edges). Results indicate that Inductive Miner consistently delivers the best trade‑off, achieving roughly 30 % faster runtimes and 20 % lower memory usage than heuristic or fuzzy miners, while maintaining comparable fitness and precision. Genetic mining attains the highest multi‑objective scores but requires up to five times longer computation, highlighting a classic speed‑accuracy trade‑off.

A key contribution of the paper is a four‑dimensional decision framework that maps algorithm selection to (a) data characteristics (noise level, log volume), (b) quality priorities (fitness vs. precision), (c) operational constraints (time and hardware resources), and (d) business objectives (process redesign vs. monitoring). This framework equips practitioners with a practical guide for choosing the most appropriate mining technique under specific project conditions.

The conclusion outlines future research directions, including hybrid methods that combine the structural guarantees of inductive mining with the optimization power of evolutionary algorithms, real‑time streaming log processing, and integration of process mining outputs with predictive AI models for proactive process management. The authors also stress the importance of effective visualization and user‑interface design to translate mined insights into actionable decisions across organizational levels.

Overall, the paper delivers a comprehensive, empirically grounded comparison of process mining approaches, providing both academic researchers and industry professionals with actionable insights for leveraging event‑log data to drive continuous process improvement.