End-to-end process mining on a complex multi-stage approval and reimbursement workflow — bottleneck analysis, conformance checking, SHAP-explained early warning model, survival analysis, and root cause analysis.
The system identifies high-risk cases before completion, enabling targeted intervention and reduced cycle time.
Data: 7,065 cases · 86,581 events · 51 activities · 18 months (BPI Challenge 2020, TU/e)
| Finding | Detail |
|---|---|
| 991 cases (14%) permanently stuck | Last event is Send Reminder — median duration 134d vs 63d for resolved cases |
| 17.1% travel-ordering violations | 746 Type A (departed before permit submitted), 583 Type B (departed before approval) |
| Scheduling dominates duration | 69% of case duration is voluntary employee scheduling, not admin processing |
| Deployable early warning model — k=8 events | AUC 0.810 (leakage-free, temporal CV) — triggers at Permit FINAL_APPROVED |
| Retrospective explanatory model — complete features | AUC 0.967 using full process information (not deployable — future leakage) |
| Data drift confirmed | elapsed_days feature halved from 2017Q1 → 2018Q4; k-fold overstates AUC by +0.048 |
| Temporal leakage identified & corrected | elapsed_days alone achieves AUC 0.833 — excluded from deployed model (Notebook 10) |
| Remaining time prediction at k=8 events | MAE 12.4 days — P10/P50/P90 quantile intervals, 80.8% coverage (Notebook 11) |
| Survival analysis (all 7,065 cases) | KM median 72.4d · Cox concordance 0.814 · 14% right-censored (Notebook 12) |
| Violation root cause analysis | 44.9% of cases have conformance violations (fitness < 1.0); DT AUC 0.876, XGB AUC 0.956 — department, duration, and event count are dominant drivers (Notebook 13) |
| Rejection paradox — N rejections is protective | 1-rejection cases: 98.1% violation rate. Zero-rejection violators: submitted permits (99.2%) but skipped Permit APPROVED by ADMINISTRATION (44% vs 100% in compliant cases). SHAP direction is a conditional interaction, not a marginal effect (Notebook 13, §5) |
| Lead time is not the problem | Zero-rejection violators submit permits earlier (median 35d vs 16d). Administration approves in 0 days when engaged. Violation rate increases with lead time — the issue is a routing/awareness gap, not insufficient lead time (Notebook 13, §6) |
| Cross-domain transferability — BPIC 2017 loan applications | Same pipeline on 31,509 loans · 475K events. Workflow predictability varies by domain: permit process is predictable from event 1; loan outcome remains uncertain until credit decision stage (k≈10, AUC jumps 0.647 → 0.851). Curves reflect process structure, not model superiority (Notebook 14) |
AUC rises from 0.65 at the first event to 0.81 (leakage-free) by event 8 — well before the median case length of 11 events.
Workflow predictability varies substantially across domains. The permit process becomes deterministic after a small number of events (routing signals present from event 1), whereas loan applications remain uncertain until the credit decision stage (k≈10, when offer-creation events enter the prefix). Note: feature sets, targets, and validation schemes differ — curves reflect process structure, not model superiority.
Cases with many events and few approvals drive the highest violation risk; department and budget contribute less than process behaviour.
The methodology is domain-agnostic. The same pipeline applies directly to:
- Maintenance work orders — predict overdue orders, identify bottleneck approval steps
- Procurement workflows — flag purchase requests at risk of SLA breach
- Drilling approvals — early warning on permits likely to stall at regulatory stages
- Intervention planning — estimate remaining time and survival probability for field operations
- Asset integrity workflows — root cause analysis on inspection deferral patterns
Run in order. Each notebook is self-contained and writes its outputs to outputs/.
| # | Notebook | What it does |
|---|---|---|
| 01 | 01_initial_exploration.ipynb |
Case/event stats, variant frequency, time coverage |
| 02 | 02_process_structure.ipynb |
Directly-Follows Graph, transition matrix, happy path |
| 03 | 03_process_discovery.ipynb |
Inductive Miner and Heuristics Miner Petri nets |
| 04 | 04_bottleneck_analysis.ipynb |
Waiting/service time split, stuck cases, scheduling vs admin delay |
| 05 | 05_conformance_analysis.ipynb |
Token replay fitness, travel-ordering violations by department |
| 06 | 06_predictive_analytics.ipynb |
XGBoost / RF / LogReg — AUC 0.967 on complete features (retrospective) |
| 07 | 07_shap_prefix.ipynb |
SHAP explanations + prefix-based early warning (k=1–20) |
| 08 | 08_temporal_cv.ipynb |
Temporal cross-validation, optimism bias, feature drift, concept drift |
| 09 | 09_final_report.ipynb |
6-panel dashboard, priority matrix, 5 findings, 5 recommendations |
| 10 | 10_leakage_calibration.ipynb |
Leakage audit, ablation study, calibration (Brier score, reliability diagram) |
| 11 | 11_remaining_time.ipynb |
Remaining time regression — XGBoost + quantile P10/P50/P90, temporal CV, SHAP |
| 12 | 12_survival_analysis.ipynb |
Survival analysis — Kaplan-Meier, log-rank tests, Cox PH model, risk groups |
| 13 | 13_violation_root_cause.ipynb |
Violation root cause — decision tree rules, XGBoost, SHAP, rejection paradox, lead time analysis |
| 14 | 14_transfer_bpic2017.ipynb |
Transferability — full pipeline on BPIC 2017 loan applications; cross-domain AUC comparison |
The notebooks require Python 3.13. The stack (numpy 2.x, xgboost 3.x, shap 0.52, pm4py 2.7) does not work on Python 3.12 with a standard Anaconda environment due to a numpy ABI conflict.
# Verify you have Python 3.13
python3.13 --versionIf you need to install it: https://www.python.org/downloads/
git clone https://github.com/djimrastephane/ProcessPath_AI.git
cd ProcessPath_AIpython3.13 -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activateTo run the Streamlit app only:
pip install -r requirements.txtTo run the notebooks (includes pm4py, jupyter):
pip install -r requirements-dev.txtThe raw event log is not included in this repo (33 MB binary). Download it from the 4TU Research Data repository:
https://data.4tu.nl/articles/dataset/BPI_Challenge_2020/12703980
van Dongen, Boudewijn (2020): BPI Challenge 2020: Travel Permit Data. Version 1. 4TU.ResearchData. dataset. https://doi.org/10.4121/uuid:ea03d361-a7cd-4f5e-83d8-5fbdf0362550
Place the file at:
data/raw/PermitLog.xes
Live demo: https://processpathai-fejiac7urktgbcvbbwylhd.streamlit.app
Run locally:
pip install -r requirements.txt
streamlit run app/app.py8 pages:
| Page | What it shows |
|---|---|
| Overview | Dataset KPIs, monthly volume, top 20 variants, key findings |
| Bottlenecks | 6 tabs: Activity Impact (impact score + CSV), Process Flow (graphviz DOT), Department Impact (CSV), Avoidable Delay + Cost Impact (€M estimate, Plotly waterfall, CSV), Priority Matrix (Plotly scatter, 4 quadrants), Detail |
| Conformance | Token replay fitness, Type A/B travel-ordering violations by department |
| Early Warning | AUC vs prefix length, leakage audit, calibration, live k=8 prediction with SHAP waterfall |
| Remaining Time | MAE/R² vs prefix, quantile P10/P50/P90 intervals, live regression prediction with SHAP |
| Survival Analysis | Kaplan-Meier curves, Cox PH hazard ratios, risk group curves (right-censored stuck cases included) |
| Violation Root Cause | Department/budget/seasonal drivers, decision tree rules, SHAP, rejection paradox, lead time analysis |
| Process Variants | Variant frequency table, duration vs stuck rate scatter, step-by-step path inspector |
A business labels toggle in the sidebar replaces raw event log terminology (e.g. Permit FOR_APPROVAL by SUPERVISOR) with executive-friendly labels (e.g. Supervisor Review) across all pages.
Cost model (sidebar): set cost_per_waiting_day, avg_travel_cost, and avg_salary_cost — the dashboard instantly recalculates estimated annual cost of delays, avoidable savings, and per-activity cost impact in euros. Example: 69,000 avoidable days × €180/day = €12.4M annual impact.
The dashboard answers six operational questions:
| Question | Where to find it |
|---|---|
| How many cases are delayed? | Overview → stuck cases KPI; Bottlenecks → detail tab |
| Where do delays occur? | Bottlenecks → Activity Impact tab (impact score ranking) |
| Who owns them? | Bottlenecks → Department Impact tab (total delay by team) |
| What is their business impact? | Bottlenecks → Avoidable Delay tab (€M cost estimate, per-activity cost) |
| What should management fix first? | Bottlenecks → Priority Matrix tab (Act Now / Investigate / Monitor / Ignore quadrants) |
| What savings are possible? | Bottlenecks → Avoidable Delay tab (avoidable cost KPI + savings from top bottleneck) |
CSV exports available on Activity Impact, Department Impact, Avoidable Delay, Priority Matrix, and Process Variants tables.
A production-ready prediction API wrapping all three deployed models.
Run locally:
pip install -r requirements.txt
uvicorn api.main:app --reload
# → http://127.0.0.1:8000/docs (Swagger UI, auto-generated)Docker:
docker build -t processpathai .
docker run -p 8000:8000 processpathaiEndpoints:
| Method | Path | Description |
|---|---|---|
GET |
/health |
Liveness check — confirms all three models are loaded |
GET |
/models |
Model metadata: type, AUC, MAE, concordance, feature counts |
POST |
/predict |
Full case prediction — see schema below |
Example request:
curl -X POST http://127.0.0.1:8000/predict \
-H "Content-Type: application/json" \
-d '{
"n_events": 8, "elapsed_days": 12.5,
"n_rejections": 1, "n_reminders": 0, "n_approvals": 2,
"case_start_month": 3, "case_start_dow": 1,
"has_send_reminder": false, "has_final_approved": false,
"has_any_rejection": true, "has_any_approval": true
}'Example response:
{
"early_warning": { "probability_late": 0.41, "risk_level": "MEDIUM", "model_auc": 0.81 },
"remaining_time": { "point_estimate_days": 10.0, "p10_days": 3.8, "p50_days": 10.0, "p90_days": 47.0 },
"survival": { "prob_complete_by_90d": 0.69, "median_survival_days": 65.4, "concordance_index": 0.814 }
}python -m ipykernel install --user --name python313 --display-name "Python 3.13"jupyter notebookOpen notebooks in order from the notebooks/ directory. Select kernel Python 3.13 when prompted.
for nb in notebooks/[01][0-9]_*.ipynb; do
jupyter nbconvert --to notebook --execute --inplace \
--ExecutePreprocessor.timeout=600 \
--ExecutePreprocessor.kernel_name=python313 \
"$nb"
doneEach notebook writes figures to outputs/figures/ and tables to outputs/tables/. Pre-computed outputs are already committed so you can browse results without re-running.
ProcessPath_AI/
├── notebooks/ # 14 analysis notebooks (run in order)
├── src/ # Shared loader and helper functions
│ ├── load_event_log.py
│ ├── inspect_log.py
│ └── process_summary.py
├── outputs/
│ ├── figures/ # 70+ PNG charts (pre-computed)
│ └── tables/ # 50+ CSV tables (pre-computed)
├── app/
│ ├── app.py # Streamlit decision support app (8 pages, cost model, priority matrix)
├── api/
│ └── main.py # FastAPI REST API (3 endpoints)
│ └── model/
│ ├── prefix_k8.joblib # Early warning classifier (AUC 0.810)
│ ├── remaining_time_k8.joblib # Remaining time regressor (MAE 12.4d)
│ └── survival_cox_k8.joblib # Cox PH model (concordance 0.814)
├── data/
│ └── README.md # Data download instructions
├── requirements.txt
├── requirements-dev.txt
└── main.py # CLI entry point (prints dataset summary)
| Library | Version | Purpose |
|---|---|---|
| pm4py | 2.7.22.5 | XES loading, DFG, Petri nets, conformance |
| pandas / numpy | 2.x | Data wrangling |
| scikit-learn | ≥1.3 | Preprocessing, CV, metrics |
| xgboost | 3.x | Gradient boosting classifier |
| shap | 0.52 | Feature attribution (TreeExplainer) |
| lifelines | ≥0.30 | Survival analysis (Kaplan-Meier, Cox PH) |
| matplotlib | ≥3.7 | Static figures (notebooks + app) |
| plotly | ≥5.20 | Interactive charts in app (priority matrix, cost waterfall) |
If you use this dataset, please cite:
van Dongen, Boudewijn (2020): BPI Challenge 2020: Travel Permit Data. Version 1. 4TU.ResearchData. dataset. https://doi.org/10.4121/uuid:ea03d361-a7cd-4f5e-83d8-5fbdf0362550
MIT