Outcome-Oriented Prescriptive Process Monitoring Based on Temporal Logic Patterns

This repository contains the source code of a new prescriptive process monitoring system that provides process analysts with recommendations for achieving a positive outcome of an ongoing process. The recommendations are temporal relations between the activities being executed during an ongoing case of a process. The paper describing the system is currently being under review, a preprint can be found here.

Repository Structure

• media/input contains the input logs in .csv format. Before reproducing the experiments it is necessary to download and unzip the log folder from here;
• media/output contains the numeric results regarding the performance of the prescriptive system;
• models contains the pre-trained models trained on the datasets described in the paper.
• src contains the backbone code;
• settings.py contains the main settings for the experiments as described in the paper below;
• dataset_figures.py is a Python script to extract the dataset figures and save them in a .csv file in the media/output folder;
• experiments_runner.py is the main Python script for running the experiments;
• gather_results.py is a Python script for aggregating the results of each dataset and presenting in a more understandable format.

Requirements

The following Python packages are required:

• numpy tested with version 1.25.0;
• PM4PY tested with version 2.7.4;
• sklearn tested with version 1.2.2;
• pandas tested with version 2.0.2.
• matplotlib tested with version 3.7.1;
• imbalanced-learn tested with version 0.10.1;
• https://seaborn.pydata.org/ tested with version 0.12.2.

Usage

The system has been tested with Python 3.11.4. After installing the requirements, please download this repository.

Running the code

To run the evaluation for a given (pretrained) dataset, type:

python3 experiments_runner.py --log=Production

if you want to train again your model, you need to set the --load_model option:

$ python experiments_runner.py --log=Production --load_model

The DECLARE constraints used for the encoding are grouped into five families: existence, choice, positive relations, negative relations and all. If you need to use a subset of constraints (e.g., existence and choice) use:

$ python experiments_runner.py --log=Production --decl_list="existence,choice"

You can also train a model on your own dataset my_event_log saved in the standard .csv format for event log. First of all, define you have to add the needed keys to the configuration dictionaries in the src/dataset_manager/DatasetManager.py file:

dataset = "my_event_log"
filename[dataset] = os.path.join(logs_dir, "my_event_log.csv")
case_id_col[dataset] = "CaseID_my_event_log"
activity_col[dataset] = "ActivityID_my_event_log"
resource_col[dataset] = "ResourceID_my_event_log"
timestamp_col[dataset] = "CompleteTimestamp_my_event_log"
label_col[dataset] = "labelID_my_event_log"
pos_label[dataset] = "pos_label_my_event_log"
neg_label[dataset] = "neg_label_my_event_log"

then, you just need to run

$ python3 experiments_runner.py --log=my_event_log --load_model=False

Type:

./run_experiments.sh

to run the experiments on the whole pool of datasets in parallel. This works in Unix systems, for Windows systems you need to create your .bat file or run each single dataset as shown above.

Gathering the results

After running the experiments, type:

$ python plot_time_performance.py

to aggregate the computational times of generating the recommendations for all datasets. The results will be in the file aggregated_recommendation_times.pdf in media/output/result.

Type

$ python gather_results.py

to have an aggregation of the results for each dataset. Such aggregation are found in the files in the media/output folder.

Citing

If you use our prescriptive process in your research, please use the following BibTeX entry

@article{abs-2211-04880,
  author       = {Ivan Donadello and
                  Chiara Di Francescomarino and
                  Fabrizio Maria Maggi and
                  Francesco Ricci and
                  Aladdin Shikhizada},
  title        = {Outcome-Oriented Prescriptive Process Monitoring Based on Temporal
                  Logic Patterns},
  journal      = {CoRR},
  volume       = {abs/2211.04880},
  year         = {2022}
}