pmf-liris/pmf-concurrency-models

Concurrency-aware process model forecasting: models and drawings

Weekly process models for the BPI2017, BPI2019 and Hospital Billing event logs, produced by the code of the paper Concurrency-Aware Process Model Forecasting with Causal Nets, and a drawing of each one as a workflow net. The code is at github.com/YongboYu/pmf-concurrency.

Contents

models/<tag>/<Log>_pmf/weekly_models/window_<n>_<family>.json   causal nets
models/<tag>/<Log>_pmf/structural_metrics.csv                  net sizes and AND share
drawings/<tag>/<Log>/index.csv                                 one row per model
drawings/<tag>/<Log>/<family>/window_<n>-expanded.svg          workflow net as constructed
drawings/<tag>/<Log>/<family>/window_<n>-reduced.svg           the same net after τ-reduction
drawings/<tag>/<Log>/static/static-expanded.svg, static-reduced.svg
SHA256SUMS

<tag> is the binding filter: tpat0 for tPAT = 0.0, which keeps bindings at or above the mean, and tpatm1 for tPAT = -1.0, which keeps every binding. Each evaluation window spans 7 days, and consecutive windows start one day apart.

Log	Windows
BPI2017	58
BPI2019	55
Hospital_Billing	139

Model families

• Forecast: reconstructed from Chronos-2 forecasts of the daily relation and binding counts.
• Re-mined: the same reconstruction applied to the counts observed in the window.
• Direct-mined: Fodina applied directly to the events of the window.
• Static: Fodina applied once to the training log and used for every window.

Forecast and Re-mined are included at both tPAT settings. Direct-mined and Static are included at tPAT = 0.0, as in the paper. structural_metrics.csv is the file the pipeline writes and lists every family at its setting.

The Static model is the same in every window, so it is drawn once and its index row has no window.

Browse the drawings

The Space pmf-liris/pmf-concurrency-viewer shows any drawing with pan and zoom and puts several side by side for comparison.

Each link opens a folder of drawings; click a file to view it in the browser.

Log	tPAT
BPI2017	0.0	Forecast	Re-mined	Direct-mined	Static
BPI2019	0.0	Forecast	Re-mined	Direct-mined	Static
Hospital Billing	0.0	Forecast	Re-mined	Direct-mined	Static
BPI2017	-1.0	Forecast	Re-mined
BPI2019	-1.0	Forecast	Re-mined
Hospital Billing	-1.0	Forecast	Re-mined

Reading a drawing

• Green circle: the source place. Pink double circle: the sink place.
• Rounded box: an activity, with its count in grey. ▶ and ■ are the start and end markers.
• Black bar: a silent transition that carries an input binding (IB=n, green) or an output binding (OB=n, blue).
• Number in a place: the succession count of the pair of activities it connects.
• Counts are window totals. Forecast counts below 1 keep two decimals, and counts below 0.01 are shown as <0.01.

The title of each drawing gives the log, family, window dates and tPAT. The reduced drawing keeps the node names of the expanded one. Large nets, such as BPI2019 Static, are best viewed zoomed in.

Index

Each index.csv has one row per model: log, tpat, family, window, start_date and end_date (inclusive, YYYY-MM-DD), the source model file, empty (true for a model without tasks, which is not drawn), the number of places, transitions, silent transitions and arcs before and after reduction, and files, the drawings relative to the index. The dataset viewer shows all six index files as one table.

Reproduction

The models were produced by commit 27b0172 with Chronos-2 revision 29ec3766d36d6f73f0696f85560a422f50e8498c on one NVIDIA RTX A6000 (CUDA 12.8, PyTorch 2.10.0). Forecast models depend on the device: another GPU or a CPU can give slightly different forecasts. The other families do not use Chronos-2.

The drawings were made with Graphviz 12.2.1 by scripts/render_nets.py at commit e12acc8. From a clone of the repository:

hf download pmf-liris/pmf-concurrency-models --repo-type dataset \
    --local-dir pmf-concurrency-models
uv run python scripts/render_nets.py \
    --output-base pmf-concurrency-models/models/tpat0 --out-dir drawings/tpat0
uv run python scripts/render_nets.py \
    --output-base pmf-concurrency-models/models/tpatm1 --out-dir drawings/tpatm1 \
    --families forecast re_mined

To produce the models themselves, follow REPRODUCE.md in the repository. The event logs are not included here; the code fetches them from Zenodo record 18327515.

License

CC BY 4.0. The code is MIT-licensed.