Typos in Event Logs Can Quietly Wreck Process Mining, Study Warns

A single mistyped character in a business process log sounds trivial. But a new study argues that these tiny errors can quietly dismantle one of the most important tools companies use to understand how their work actually gets done. In research published in the International Journal of Data Science and Analytics, a team led by Marco Comuzzi of Ulsan National Institute of Science and Technology, together with colleagues at Kunsan National University and Politecnico di Milano, systematically injected four kinds of label errors into event logs and measured the damage across the three core tasks of process mining: process discovery, conformance checking, and predictive process monitoring. Their findings upend the intuition that all data errors deserve equal attention, showing instead that the cost of a dirty label depends dramatically on what you plan to do with it.

Process mining is the discipline of reconstructing business processes from the digital footprints left in information systems. Every event log must record at least three things: a case identifier that ties events to a single process instance, a timestamp, and an activity label naming the step that occurred. Those activity labels are the backbone of nearly every process mining technique, which makes them a tempting target for anyone asking the classic data science question: how much does garbage in really cost? The researchers framed their work around a simple research question: how do errors on activity labels affect the performance of discovery, conformance checking, and predictive monitoring?

To answer it, the team leaned on a well-established taxonomy of event log imperfection patterns developed by Suriadi and colleagues, focusing on the four that corrupt categorical labels. Distorted labels arise from typos, simulated in the experiments through five operations such as deleting a character, inserting a random one, swapping two characters, substituting a neighboring keyboard key, or flipping case. Polluted labels occur when extra information, such as a resource identifier or a random string, gets glued onto the activity name. Homonymous labels use one name for several genuinely different steps, while synonymous labels describe the same step with different names, a common artifact of merging logs from multiple systems. Using scripts built on the FLAWD language, the researchers injected each error type separately into clean logs at controlled error ratios, from as low as 2 percent of events up to 50 percent for the prediction experiments.

The experimental design cleverly mirrors real-world scenarios. For process discovery, models were built from dirty logs and then tested against clean logs, simulating a hospital that mines a process from error-prone records at one site and applies the resulting model at another site with clean data. For conformance checking, the direction was reversed: models discovered from clean logs were used to replay dirty logs, reflecting organizations that hold a trusted reference model and audit live, imperfect data. For predictive monitoring, long short-term memory neural networks with two layers of 128 nodes were trained on noisy logs and evaluated on clean test data across three tasks: next activity prediction, outcome prediction, and remaining time prediction. The experiments drew on widely used real-life logs from the Business Process Intelligence Challenges of 2011 and 2015, two synthetic datasets that allowed controlled injection of homonyms and synonyms, and a real log from the Italian judicial system.

The starkest result came from process discovery. Fitness, which measures how well a discovered model can replay the clean log, barely budged, because most traces in a dirty log remain clean. Precision, however, collapsed. Error patterns that inject many new labels, namely random distortions and random pollution, drove precision close to zero with as little as 2 percent of events affected, because the discovery algorithm responds to the flood of novel trace variants by producing wildly permissive, underfitting models. Patterns that introduce only a handful of new labels, such as homonyms and synonyms, were far gentler. The practical implication is uncomfortable: cleaning for discovery must be nearly perfect to help, since even a small residue of random errors can destroy model precision, and analysts rarely know the true error rate in advance.

Conformance checking told a much more forgiving story. When a clean reference model was replayed against dirty logs, fitness declined only modestly, by roughly 10 percent in the worst case at a 10 percent error ratio, while precision stayed essentially constant. In many real settings, the authors argue, such degradation may be negligible for decision-making. This asymmetry carries a clear managerial message: it is far riskier to discover a process model from dirty data than to check dirty data against a trustworthy model. The quality of the reference process model, not the cleanliness of every incoming log, is the true fulcrum of reliable process mining.

Predictive process monitoring showed the sharpest differentiation among tasks. Next activity prediction, a multi-class classification problem in which label errors directly corrupt the target classes, suffered massive degradation that scaled with the number of new labels each error pattern introduced. Random pollution was worst, typos came next, and systematic pollution was milder, with performance at a 30 percent error ratio falling to roughly 30 percent of the clean-data baseline. Outcome prediction and remaining time prediction, by contrast, proved remarkably robust, with errors in mean absolute error generally staying within 10 percent of baseline. The researchers attribute this to the models leaning on numerical temporal features, such as time since case start and time since the previous event, rather than the exact wording of activity labels. In a striking twist under mixed error conditions, the LSTM networks eventually learned to discount the noisy categorical features entirely and rely on the unpolluted time series, stabilizing their predictions.

A real-world case study in the Italian civil judicial system grounded these findings outside the laboratory. The telematic case management system SICID defines more than 1,400 event types whose abstract and textual descriptions frequently diverge, creating naturally polluted and ambiguous labels. Training a remaining time predictor on 24,465 finished cases yielded a mean absolute error of 212 days with standardized clean codes versus 269 days with polluted labels, a difference of nearly two months in predicting trial durations. For a justice system under constitutional pressure to deliver timely hearings, that gap is not an academic footnote; it directly affects resource planning and the fairness of proceedings.

Perhaps the most counter-intuitive guidance from the study concerns cleaning priorities. Because errors that multiply the number of distinct activity labels, especially random typos and random pollution, cause the most damage, the authors recommend prioritizing cheap, automated fixes for random errors over expensive expert-driven efforts to resolve synonyms and homonyms, which introduce few new labels and correspondingly little harm. Techniques borrowed from tabular data cleaning, such as the Sorted Neighborhood deduplication method, edit distance metrics, word embeddings, transformer-based semantic similarity, and domain ontologies, can all be deployed to detect and repair label errors. Conformance checking itself can even serve as a diagnostic, with alignment discrepancies pinpointing which labels need targeted repair.

The study is candid about its limits: it relies on the Heuristic Miner discovery algorithm and LSTM-based predictors, restricts itself to imperative process models, and uses random uniform error injection that may not capture every real-world failure mode. Yet the authors see no strong theoretical reason why other algorithms would behave fundamentally differently. What emerges is a practical knowledge base for anyone designing process mining pipelines: know your downstream task before you clean, spend your data quality budget where new labels multiply, and treat a trusted process model as your best defense against dirty data. In an era when organizations increasingly hand their operations to data-driven analysis, this work is a reminder that the humble activity label, the word naming each step of a process, can be the difference between insight and illusion.

Subject of Research: The impact of activity label errors in event logs on process mining tasks such as process discovery, conformance checking, and predictive process monitoring

Article Title: On the impact of low-quality activity labels on data-driven process analytics

Article References: Comuzzi, M., Kim, S., Ko, J., Cappiello, C., Salamov, M., & Pernici, B. (2026). On the impact of low-quality activity labels on data-driven process analytics. International Journal of Data Science and Analytics, 22(1), Article 340. https://doi.org/10.1007/s41060-026-01334-z

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01334-z

Keywords: process mining, event logs, data quality, activity labels, process discovery, conformance checking, predictive process monitoring, data cleaning, LSTM, machine learning, imperfection patterns, business process management