Implementation of NEFCLASS in python

This repository implements the neuro-fuzzy NEFCLASS model in python

Requisites

First install anaconda onto your system.

conda create -n nefclass python==3.6.10
conda install pandas scikit-learn

Bulk of the code is done in pure numpy. pandas is used for loading csv files, and scikit-learn is used for creating kfolds cross validation sets.

Data folder

Git clone this repository and put the wine dataset or the iris dataset in data/. If you want to extend to your own dataset, you can add the data loading function to data_loading.py and then add the appropriate function to main.py

The data folder should look like this

data/Iris.csv
data/wine/winequality-red.csv
data/wine/winequality-white.csv
data/wine/winequality.names

Commands

usage: main.py [-h] [--dataset DATASET] [--sigma SIGMA]
               [--num_epoch NUM_EPOCH] [--num_sets NUM_SETS] [--kmax KMAX]
               [--rule_learning RULE_LEARNING] [--cv] [--kfold KFOLD] [-v]
               [--mf MF]

NEFCLASS

optional arguments:
  -h, --help            show this help message and exit
  --dataset DATASET     dataset to load
  --sigma SIGMA         learning rate
  --num_epoch NUM_EPOCH
                        number of epoch for fuzzy set learning
  --num_sets NUM_SETS   number of fuzzy sets
  --kmax KMAX           maximum number of rules
  --rule_learning RULE_LEARNING
                        rule learning method to use. Default is the original
                        implementation. Use any other strings as input to
                        perform best per class.
  --cv                  do 10 fold cross validation?
  --kfold KFOLD         number of k fold
  -v                    verbosity
  --mf MF               membership function to use. Default: tri. Options:
                        gaussian, semicircle

Example commands

To train on iris dataset on default arguments with verbose printing
python main.py -v

To perform cross validation on iris dataset
python main.py --cv

To train on wine dataset, with best per class rule learning, and learning rate of 0.1;
python main.py -v --dataset wine --rule_learning bestperclass --sigma 0.1

Acknowledgements

This project was done as a part of the fulfilment of a module at Nanyang Technological University