This repo contains scripts to train models for Nirdizati predictive process monitoring engine. The following instructions are for the standalone use. For use with Apromore, please refer to these instructions.

Requirements

Tested with Python 3.5 and Python 3.6. Install the necessary packages with

pip install -r requirements.txt

Training predictive models

export PYTHONPATH=$PYTHONPATH:/wherever/you/keep/nirdizati-training-backend
cd core/
python train.py training-config-ID 

• training-config-ID - JSON file (without extension) that contains training configuration (see below), must be placed under core/training_params/ directory

Example:

python train.py myconfig_remtime

This script assumes that you have a training configuration file core/training_params/myconfig_remtime.json with the following structure:

Training configuration structure

{
  "target": {
    "bucketing_type": {
      "encoding_type": {
        "learner_type": {
          "learner_param1": "value1",
          "learner_param2": "value2",
          ...
        }
      }
    }
  },
  "ui_data": {
    "log_file": "/wherever/you/keep/your/log.csv"
  }
}

• bucketing_type - zero, cluster, state or prefix. Use tooltips in the Nirdizati tool (advanced training mode) for short explanation
• encoding_type - agg, laststate, index or combined
• learner_type - rf for random forest, gbm for gradient boosting, dt for decision tree or xgb for extreme gradient boosting
• learner_param's - most important hyperparameters for each learner, taken from sklearn lib: n_estimators and max_features for rf; n_estimators, max_features and learning_rate for gbm; max_features and max_depth for dt; n_estimators, max_depth, learning_rate, colsample_bytree and subsample for xgb
• target - variable that you want to predict (see below). The prediction problem type (classification or regression) is determined automatically based on the number of unique levels of a target variable and whether or not it can be parsed as a numeric series.

Example of a training configuration file

What can be predicted?

• Remaining cycle time. Use remtime keyword as a target argument for train.py
• Binary case outcome based on the expected case duration (whether case duration will exceed a specified threshold). Use a positive threshold value for target or "-1" if you want the labeling to be based on the median case duration.
• Next activity to be executed. Use next for target
• Any static, i.e. case, attribute that is already available in the log as a column. In this case, target is the name of the corresponding column.

Output of the training script:

• Fitted model - pkl/
• Validation results by prefix length - results/validation/
• Detailed validation results - results/detailed/
• Data on feature importance - results/feature_importance/

How to choose default training parameters?

Bucketing - No bucketing (zero)

Encoding - Frequency (agg)

Predictor - XGBoost

Default hyperparameters for XGBoost predictor:

• Random forest: Number of estimators 300, max_features 0.5
• Gradient boosting: Number of estimators 300, max_features 0.5, learning rate 0.1
• Decision tree: max_features 0.8, max_depth 5
• XGBoost: Number of estimators 300, learning rate 0.04, subsample row ration 0.7, subsample column ratio 0.7, max_depth 5

Test for an ongoing case

export PYTHONPATH=$PYTHONPATH:/wherever/you/keep/nirdizati-training-backend
cd core/
python predict_trace.py path_to_single_test_prefix.json path_to_pickle_model_filename 

Example:

python predict_trace.py ../logdata/1170931347.json ../pkl/bpi17_sample_myconfig_remtime.pkl

The output should be printed to stdout, for example:

[{"remtime":15.9565439224}]