Inferring A Multi-Perspective Likelihood Graph from Black-Box Next Event Predictors

This repository contains a prototype implementation of the method described in the paper Inferring A Multi-Perspective Likelihood Graph from Black-Box Next Event Predictors.

Abstract

Deep learning models for next event prediction in predictive process monitoring have shown significant performance improvements over conventional methods. However, they are often criticized for being black-box models. Without allowing analysts to understand what such models have learned, it is difficult to establish trust in their abilities. In this work, we propose a technique to infer a likelihood graph from a next event predictor to capture and visualize its behavior. Our approach first generates complete cases, including event attributes, using the next event predictor. From this set of cases, a multi-perspective likelihood graph is inferred. Including event attributes in the graph allows to better identify decision and branching points within a process. The evaluation shows the inferred graphs generalize beyond the event log and achieve high F-scores and small likelihood deviations. We conclude black-box NEPs can be used to generate correct cases even in the presence of anomalies in the event log. As a result, our visualization technique, which represents exactly this set of cases, shows what the NEP has learned, thus mitigating one of their biggest criticisms.

Example Likelihood Graphs

The following likelihood graph was generated using our approach. Red nodes represent activities, yellow nodes the event attribute user. More examples can be found in /examples.

Jupyter Lab Notebooks

Check the notebooks directory for example Jupyter Lab Notebooks. Follow the setup guide and make sure to install the ipywidgets extension (https://ipywidgets.readthedocs.io/en/latest/user_install.html) for Jupyter Lab.

Setup

The easiest way to setup an environment is to use Miniconda.

Using Miniconda

1. Install Miniconda.
2. After setting up miniconda you can make use of the conda command in your command line (Powershell, CMD, Bash).
3. Set up a dedicated environment for this project by running conda env create -f environment.yml
   • This will setup a virtual conda environment with all necessary dependencies.
   • Installs the CPU only version of Tensorflow.
4. Depending on your operating system you can activate the virtual environment with conda activate neplg.
5. If you want to quickly install the neplg package, run pip install -e . inside the root directory.
6. Now you can start the notebook server by jupyter lab notebooks.

Note: To use the graph plotting methods, you will have to install Graphviz.

IDE

We recommend you use PyCharm Community Edition as your IDE.