Predictive Monitoring of business processes

This repository contains the code from my master's thesis "Feature extraction from event logs for predictive monitoring of business processes", which is included as thesis_fborchert.pdf.

There are 2 subprojects:

• feature_extraction contains the implementation of tree-based trace profiles and general classes for transforming event log cases to WEKA instances using a trace profile
• process_prediction contains the experimental setup with different trace profiles in a binary classification and an outlier detection setting

How to run the experiments for your own logs

Requirements:

• Scala 2.11
• sbt

Configuration

You need to edit the file process_prediction/config.properties to specifiy the parameters for your data.

• title - the title of your experiments for the result files, e.g. title = My Experiment
• type - the type of algorithm you want to use, valid values are type=randomforest and type=outlier, but extensions should be straightforward to implement
• training- the folder containing the training data in XES format
• test - the folder containing the hold out test data in XES format
• ingore-attributes - a comma-separated list of case attributes (by regular expressions) that should be ignored, e.g. ignore-attributes=year,applicant,payment_granted[^0].*,payment_actual[^0].*,cutting[^0].*
• prefix - the full qualified class name of a Scala class that maps traces to prefixes, e.g. prefix=input.ExamplePrefix (see the process_prediction\src\main\scala\input folder for an example)
• output - the full qualified class name of a Scala class that maps traces to labels. If type=outlier, this label is used for evaluation (see the process_prediction\src\main\scala\input folder for an example)
• IM-noise - the noise threshhold for the Inductive Miner used to construct tree-based trace profiles, e.g. IM-noise=0.8. The parameter is optional and chosen via cross-validation if not supplied
• nu- the nu parameter of the OC-SVM used for outlier detection, e.g. nu=0.05. This parameter is optional and set to 0.05 if absent and type=outlier has been set

Run it

Go to the root folder of this project and run the following commands:

• sbt (opens the sbt console)
• compile
• project process_prediction
• run

When you properly configured the experimental setup, this should run a series of experiments and place results in a folder named results/<title>

Supplying different trace profiles

You can change the implementation of the getConfigs() methods in:

• process_prediction/src/main/scala/de/huberlin/informatik/promi/experiments/BinaryClassification.scala
• process_prediction/src/main/scala/de/huberlin/informatik/promi/experiments/outlier/Outlier.scala

to modify the experimental setup, edit the trace profiles used or supply your own.