TJNOV / pm-prediction

A simple evaluation pipeline of predictive business process monitoring techniques

Overview

This platform supports the evaluation of novel (instance-level) predictive business process monitoring techniques. It supports

• importing event logs,
• generating basic features (e.g., activity history, service time, resource history),
• plugging in model architectures,
• training models, and
• evaluating models.

How to

1. Clone repository

   • git clone

2. Install dependencies (Python 3.8.8)

   • pip install -r requirements.txt

3. How to:

   1. You can find example scripts at ./sample_scripts/
   • Run the script with the following flags
     • --task: "next_activity",
     • --contextual_info: "True" if you want to exploit contextual features, "False" otherwise (default: False),
     • --transition: "True" if the dataset contains transition information, "False" otherwise (default: False),
     • --learning_rate: learning rate of optimization algorithm (default: 0.002),
     • --num_folds: number of folds in validation (default: 10),
     • --batch_size: training batch size of models (default: 256),
     • --data_dir: directory where the dataset is located,
     • --data_set: name of the dataset,
     • --checkpoint: directory where the models are saved (default: "./checkpoints/"),
     • --control_flow_p: "True" if you want to use control-flows as features, "False" otherwise (default: True),
     • --time_p: "True" if you want to use time as features, "False" otherwise (default: True),
     • --resource_p: "True" if you want to use resource as features, "False" otherwise (default: False),
     • --data_p: "True" if you want to use data as features, "False" otherwise (default: False),
     • --cross_num: number of cross validations
     • --result_dir: directory where the experimental results are saved
   • e.g.,
     • python sample_scripts/next_activity_with_neural_network.py --task "next_activity" --contextual_info False --num_epochs 10 --batch_size 256 --data_dir "./samples/" --data_set "bpi_2013_closed_problems.xes.gz_Train_All" --test_data_dir "./samples/" --test_data_set "bpi_2013_closed_problems.xes.gz_Test" --checkpoint_dir "./checkpoints/" --control_flow_p True --time_p True --resource_p False --data_p False --transition False --cross_number 5 --result_dir "./result/"
     • python sample_scripts/next_activity_with_xgboost.py --task "next_activity" --contextual_info False --data_dir "./samples/" --data_set "bpi_2013_closed_problems.xes.gz_Train_All" --test_data_dir "./samples/" --test_data_set "bpi_2013_closed_problems.xes.gz_Test" --checkpoint_dir "./checkpoints/" --control_flow_p True --time_p True --resource_p False --data_p False --transition False --cross_number 5 --result_dir "./result/"

Requirements

• We use the datetime format: "%Y.%m.%d %H:%M".
• We assume the input csv file contains the columns named after the xes elements, e.g., concept:name
• We assume csv format dataset where 'case:concept:name' denotes the caseid, 'concept:name' the activity,'org:resource' the resource, 'time:timestamp' the complete timestamp, 'event_@' the event attribute, and 'case_@' the case attribute.