DupliMend: Online Detection and Refinement of Imprecise Activity Labels

Source code and additional resources for the paper "DupliMend: Online Detection and Refinement of Imprecise Activity Labels" by Savandi Kalukapuge, Andrzej Janusz, and Moe Thandar Wynn (CAiSE 2026).

Key Contributions

1. A streaming, unsupervised framework that detects and refines homonymous activity labels on-the-fly without prior specification, complete traces, or full-log analysis.

2. Activity-specific sparse denoising autoencoders that learn multi-perspective representations and enable dynamic splitting and merging of label variants as behaviour evolves through online machine learning, specifically, online clustering.

3. A drift-aware continual learning mechanism combining ADWIN, cluster regularisation, and centroid memory replay to maintain stability and prevent forgetting under evolving process event streams.

Approach Overview

Quick Start

Run DupliMend on the included I-PALIA dataset and reproduce evaluation metrics.

Step 1: Clone and Install

git clone https://github.com/Savandi/DupliMend.git
cd DupliMend
python -m venv venv
source venv/bin/activate  # Linux/Mac (or: venv\Scripts\activate on Windows)
pip install -r requirements.txt

Step 2: Run DupliMend

The repository includes the I-PALIA synthetic dataset:

• Input: src/synthetic_logs/ipalia.csv (~8,400 events)
• Ground truth: src/synthetic_logs/ipalia_groundtruth.csv

This dataset contains a homonymous activity label "A" that should be refined into three variants: A_start, A_middle, and A_end.

python main.py

This will process events in streaming fashion, train activity-specific autoencoders online, cluster events, and output results to run_output/tracking_YYYYMMDD_HHMMSS/.

Step 3: Evaluate Results

After main.py completes, evaluate clustering quality against ground truth:

# Set the tracking directory (replace with actual output folder name)
export TRACKING_DIR=run_output/tracking_YYYYMMDD_HHMMSS  # Linux/Mac
# $env:TRACKING_DIR="run_output\tracking_YYYYMMDD_HHMMSS"  # Windows PowerShell

cd src/evaluation_scripts
python evaluate_single_test_file.py

Results are saved to evaluation_results/.

---

Evaluation Metrics

Clustering Quality (Supervised):

• ARI (Adjusted Rand Index) - alignment with ground truth, adjusted for chance
• NMI (Normalised Mutual Information) - information-theoretic cluster-label agreement
• Expected Entropy (Cluster) - entropy of ground truth labels within clusters
• Expected Entropy (Label) - entropy of cluster assignments per ground truth label

Clustering Quality (Unsupervised):

• Silhouette Score - cohesion and separation measure

Discovered Model Quality:

• Log Fitness (Recall) - fraction of log behaviour reproducible by the model
• Log Precision - fraction of model behaviour observed in the log
• F-score - harmonic mean of Precision and Fitness

Streaming Performance:

• Convergence speed at intermediate checkpoints (25%, 50%, 75%, 100%)
• Throughput (events/second), Latency (ms/event), Peak Memory (MB)

---

Key Experimental Results

Clustering quality (ARI) on synthetic PESs

Discovered model precision across all PESs

Statistical comparison of precision (Friedman + Nemenyi)

• Using the Friedman test with Nemenyi post-hoc analysis (α = 0.05), DupliMend achieves the best mean rank (1.43) for Precision and is the only method that differs significantly from the unrefined baseline.

Precision convergence over the stream (real-life PESs)

Full experimental results: Dropbox Link

Optimised hyperparameters from Bayesian optimisation: src/bayesian_optimization/

---

Datasets

Synthetic Process Event Streams

Dataset	Events	Cases	Activities	Description
DuplicatedTasks	~1,000/log	1,000/log	varies	1,295 synthetic logs from 4TU with injected homonyms
I-PALIA	8,415	990	8 (10 GT)	Healthcare workflow with duplicated activity "A" (accessible in the synthetic logs folder)
DocReview	158,855	20,090	10 (11 GT)	Document review workflow with multi-attribute events (accessible in the synthetic logs folder)

Real-Life Process Event Streams

These event logs are obtained from 4TU.ResearchData, the standard repository for publicly available process mining datasets.

PES	Events	Cases	Activities	Avg Case Length	Publicly Available Link
BPIC2012	262,200	13,087	36	20	BPIC Challenge 2012
BPIC2013C	6,660	1,487	7	4	BPI Challenge 2013, closed problems
BPIC2013I	65,533	7,554	13	9	BPI Challenge 2013, incidents
BPIC2013O	2,351	819	5	3	BPI Challenge 2013, open problems
BPIC2017	1,202,267	31,509	26	38	BPI Challenge 2017
Road Fines	561,470	150,370	11	4	Road Traffic Fine Management Process
Env. Permits	8,577	1,434	27	6	Receipt phase of an environmental permit application process (WABO), CoSeLoG project

---

Baseline Methods

DupliMend is compared against two established offline methods:

Method	Reference	Algorithm
lblrefine	Lu et al. (BPM 2016)	Connected-components on event graph
lblsplit	van Zelst et al. (BPM 2023)	Leiden community detection with variant compression

---

Results Summary

Streaming Performance

PES	Latency (ms/ev)	Peak Memory (MB)	Throughput (ev/s)
DuplicatedTasks	19.10	50.24	~52
I-PALIA	16.42	36.72	~61
BPIC2012	49.97	98.52	~20
BPIC2017	42.80	82.36	~23
Road Fines	33.50	65.18	~30

Clustering Quality (Synthetic PESs)

Dataset	ARI	NMI	EE Cluster	EE Label
DuplicatedTasks (avg)	0.96	0.94	0.08	0.06
I-PALIA	0.92	0.89	0.12	0.10
DocReview	0.94	0.91	0.09	0.08

Process Model Quality Improvement

PES	Method	Fitness	Precision	F-score
Road Fines	Unrefined	0.85	0.62	0.72
Road Fines	DupliMend	0.88	0.78	0.83
BPIC2017	Unrefined	0.79	0.58	0.67
BPIC2017	DupliMend	0.82	0.71	0.76

---

Acknowledgments

This work was conducted at the Queensland University of Technology (QUT), School of Information Systems in the Faculty of Science and Centre for Data Science.

References

1. Lu, X., Fahland, D., van den Biggelaar, F. J. H. M., & van der Aalst, W. M. P. (2016). Handling duplicated tasks in process discovery by refining event labels. In BPM 2016 (LNCS 9850, pp. 90–107). Springer. DOI

2. van Zelst, S. J., Tai, J., Langenberg, M., & Lu, X. (2023). Context-based activity label-splitting. In BPM 2023 (LNCS 14159, pp. 232–248). Springer. DOI