README

D-LFIT is a software that learns propositional logic programs from tabular data, interpretation transition data, and first-order feature data generated from propositionalization methods such as BCP. The corresponding academic paper appears in @MLJ.

This code includes three main part: test D-LFIT on incomplete dataset, testing D-LFIT on incomplete dataset and testing the accuracy of generated symbolic logic program by D-LFIT. Remarkably, this code is finish based on the source code of LFIT by Tony Ribeiro (GitHub address: https://github.com/Tony-sama/pylfit). The python version: 3.0.0+ The Tensorflow version: 2.0+

Experiments

Here are some sample commands for test D-LFIT.

0. Run the following file in the path of 'D-LFIT/code/'.

1. Generate the D-LIFT readable data:

• For Boolean Networks dataset:

  python generate_traindata_Booleannetworks.py -file_name fission

  (-file_name could be: fission, mam, ara, budding)

• For relational datasets:

  python generate_traindata_relational.py

  (Generating datasets for three relational dataset: 'amine','uw-cse','mutagenesis')

2. Testing the performance of D-LFIT:

   Code formate is:

   python main.py datasets_name datasets_name incomplete? mislabelled?

   Some super-parameters such as learning rate and training times can be set in source/train_fol.py for meta-learner and source/interpretor_fol.py for interpretation learner.

• For example for Boolean datasets:

  • python main.py fission fission 1 0, execute D-LFIT on fission incomplete datasets
  • python main.py budding budding 0 1, execute D-LFIT on budding fuzzy datasets

• For example for relational datasets:

  relation dataset name and its target predicate name: 'amine: great_ne','uw-cse: uwcse1i','mutagenesis: active'

  • python main.py amine great_ne 1 0, execute D-LFIT on amine incomplete datasets
  • python main.py mutagenesis active 0 1, execute D-LFIT on mutagenesis mislabelled datasets

Citing

@article{GaoWCI22,
  author       = {Kun Gao and
                  Hanpin Wang and
                  Yongzhi Cao and
                  Katsumi Inoue},
  title        = {Learning from interpretation transition using differentiable logic
                  programming semantics},
  journal      = {Maching Learning},
  volume       = {111},
  number       = {1},
  pages        = {123--145},
  year         = {2022},
}