Probabilistic Multi-Dimensional Classification

This repository contains the code for the paper Probabilistic Multi-Dimensional Classification (UAI 2023).

Authors: Vu-Linh Nguyen*, Yang Yang*, Cassio de Campos

TL;DR: In this paper, we present a first attempt to learn probabilistic multi-dimensional classifiers which are interpretable, accurate, scalable and capable of handling mixed data.

Requirements

The required packages are listed in requirements.txt.

Data

Please check the experiment section of our paper for a collection of datasets used in the experiments.

Usage

To replicate experiments, check the scripts in the experiments folder.

To run experiments on tabular data sets with only continuous features, you can call

python cont.py Rf1 --base lr --palim 3 --n-chains 10 --no-plot --n-folds 10

usage: cont.py [-h] [--base {lr,nb}] [--palim PALIM] [--n-chains N_CHAINS] [--plot | --no-plot] [--n-folds N_FOLDS] [--output OUTPUT]
               {Edm,Jura,Enb,Voice,Song,Flickr,Fera,WQplants,WQanimals,Rf1,Pain,Disfa,WaterQuality,Oes97,Oes10,Scm20d,Scm1d}

Tabular Continuous Features Experiments

positional arguments:
  {Edm,Jura,Enb,Voice,Song,Flickr,Fera,WQplants,WQanimals,Rf1,Pain,Disfa,WaterQuality,Oes97,Oes10,Scm20d,Scm1d}
                        Dataset

options:
  -h, --help            show this help message and exit
  --base {lr,nb}        Base classifier
  --palim PALIM         The maximum number of parents for each node
  --n-chains N_CHAINS   Number of randomly generated classifier chains
  --plot, --no-plot     Whether or not to plot the BN structure
  --n-folds N_FOLDS     Number of folds
  --output OUTPUT       Output path

To run experiments on tabular data sets with both continuous and discrete features, when using a logistic regression classifier as the base learner, you can call

python disc_lr.py Adult --palim 3 --disclim 3 --n-chains 10 --no-plot --n-folds 10

usage: disc_lr.py [-h] [--palim PALIM] [--disclim DISCLIM] [--n-chains N_CHAINS] [--plot | --no-plot] [--n-folds N_FOLDS] [--output OUTPUT] {Adult,Default,Thyroid}

Mixed Data Experiments with Logistic Regression

positional arguments:
  {Adult,Default,Thyroid}
                        Dataset

options:
  -h, --help            show this help message and exit
  --palim PALIM         The maximum number of parents for each node
  --disclim DISCLIM     The maximum number of discrete features for each node
  --n-chains N_CHAINS   Number of randomly generated classifier chains
  --plot, --no-plot     Whether or not to plot the BN structure
  --n-folds N_FOLDS     Number of folds
  --output OUTPUT       Output path

To run experiments on tabular data sets with both continuous and discrete features, when using a naive bayes classifier as the base learner, you can call

python3 disc_nb.py Adult --palim 3 --disclim 3 --n-chains 10 --no-plot --n-folds 10

usage: disc_nb.py [-h] [--palim PALIM] [--disclim DISCLIM] [--n-chains N_CHAINS] [--plot | --no-plot] [--n-folds N_FOLDS] [--output OUTPUT] {Adult,Default,Thyroid}

Mixed Data Experiments with Naive Bayes

positional arguments:
  {Adult,Default,Thyroid}
                        Dataset

options:
  -h, --help            show this help message and exit
  --palim PALIM         The maximum number of parents for each node
  --disclim DISCLIM     The maximum number of discrete features for each node
  --n-chains N_CHAINS   Number of randomly generated classifier chains
  --plot, --no-plot     Whether or not to plot the BN structure
  --n-folds N_FOLDS     Number of folds
  --output OUTPUT       Output path

To run experiments on image data sets, you can call

python3 img.py voc2007 --base resnet18 --batch-size 128 --lr 0.01 --n-epochs 20 --palim 3 --no-plot --mixup --calibrate

usage: img.py [-h] [--base {resnet18}] [--batch-size BATCH_SIZE] [--lr LR] [--n-epochs N_EPOCHS] [--palim PALIM] [--plot | --no-plot] [--mixup | --no-mixup]
              [--calibrate | --no-calibrate] [--output OUTPUT]
              {voc2007}

Image Experiments

positional arguments:
  {voc2007}             Dataset

options:
  -h, --help            show this help message and exit
  --base {resnet18}     Base classifier
  --batch-size BATCH_SIZE
                        Batch size
  --lr LR               Learning rate
  --n-epochs N_EPOCHS   Number of epochs
  --palim PALIM         The maximum number of parents for each node
  --plot, --no-plot     Whether or not to plot the BN structure
  --mixup, --no-mixup   Whether or not to use mixup training
  --calibrate, --no-calibrate
                        Whether or not to calibrate
  --output OUTPUT       Output path