Description

This repository provides an implementation of the CILP++ system from [1]. It contains a copy of Aleph (obtained from the The Aleph Manual), that is used for the bottom clause propositionalization in the training pipeline.

It also includes an implementation of TREPAN [2] originally developed by Kester Jarvis and Artur d'Avila Garcez for rule extraction from the trained neural network.

The included datasets are:

1. Mutagenesis, Alzheimers from here
2. Trains, IMDb from here

Instructions

Requirements:

1. Ubuntu, Debian or similar
2. Anaconda (or Miniconda)
3. SWI Prolog
4. The required GPU drivers (optional)

To get the code and setup the environment, run:

git clone https://github.com/vakker/CILP.git
cd CILP
conda env create -f environment.yml
conda activate cilp

To run the training:

python run.py ...

The following arguments are available:

--log-dir <log-dir>   # e.g. logs
--data-dir <data-dir> # e.g. datasets/muta/muta188
--max-epochs <max-epochs>
--n-splits <n-splits>
--no-cache            # don't get data from cache instead run BCP again
--use-gpu             # use GPU for MLP
--trepan              # run a single train/val split and then TREPAN instead of cross-val
--dedup               # keep only unique data samples

To plot the training curves:

python plot.py ...

With arguments:

--log-file <log-file>     # e.g. logs/7992926c.npz (generated during training)
--param-file <param-file> # e.g. logs/params.json (also generated during training)
--max-epochs <max-epochs> # limit the number of epochs for plotting

Notes: The IMDb dataset creates a large number of features which makes TREPAN practically unusable. The Alzheimer's datasets finish a single TREPAN run within 10 min, MUTAG takes approx 1.9 hours, and IMDb was killed after 3 days.

References

[1] França, Manoel VM, Gerson Zaverucha, and Artur S. d’Avila Garcez. "Fast relational learning using bottom clause propositionalization with artificial neural networks." Machine learning 94.1 (2014): 81-104.

[2] Craven, Mark, and Jude W. Shavlik. "Extracting tree-structured representations of trained networks." Advances in neural information processing systems. 1996.