Environment Setup

OCTAVE

migration to octave

install octave with homebrew (macos):

brew install octave

or apt (linux):

apt install octave
apt install liboctave-dev

to run octave:

octave-cli -qf

install additional packages

pkg install -forge io
pkg install -forge statistics
pkg install -forge linear-algebra

install jupyter kernel in a python virtualenv

pip install virtualenv
pip install virtualenvwrapper
mkvirtualenv wmtl
pip install octave_kernel
python -m octave_kernel install --user

install code

git clone https://github.com/johnnytorres/wmtl
git submodule init 
git submodule update
git pull --recurse-submodules

AUC roc implementation

https://towardsdatascience.com/roc-curve-and-auc-from-scratch-in-numpy-visualized-2612bb9459ab

initial configuration of the project

MALSAR

only the first time install MALSAR library for baseline methods

git submodule add https://github.com/jiayuzhou/MALSAR

OPTUNITY

only at the project creation, installed modified version of optunity https://optunity.readthedocs.io/en/latest/user/installation.html#install-octave to support multitask datasets only works with python2

git submodule add https://github.com/johnnytorres/optunity.git

jsonlab

install jsonlab library to save json results

git submodule add https://github.com/fangq/jsonlab.git

Octave config

add paths for the installed libraries, change the path if it's in a different folder

addpath(genpath(make_absolute_filename('MALSAR/MALSAR/functions'))); 
addpath(genpath(make_absolute_filename('MALSAR/MALSAR/utils'))); 
addpath(genpath(make_absolute_filename('optunity-multitask/wrappers/octave/optunity'))); 
addpath(make_absolute_filename('jsonlab'));
savepath()

issues with file created in Windows OS

when trying to run from bash ./main.m it raises an error : "'usr/local/bin/octave-cli: invalid option -- '..."

verify line ending cat -e main.m

to run from command line with octave change line ending with: perl -pi -e 's/\r\n/\n/' main.m

to run with matlab change line ending with: perl -pi -e 's/\n/\r\n/' main.m

segmentation fault issue

dataset_path="/root/data/multitasktest/school/school.mat" results_path="/root/data/multitasktest/school/results/mtreg_caso/test" load (dataset_path) test_size=0.3 %pkg load statistics <-- the culprit ??? addpath('utils/'); [X_train, Y_train, X_test, Y_test] = datasplit(X, Y, test_size); solver="grid_small"; pars=struct('rho1', 100, 'rho2', 100, 'k', 1); Least_CASO(X, Y, pars.rho1, pars.rho2, pars.k, pars);

DATASET

TODO

./scripts/data.sh 20news

EXPERIMENTS

regression task experiments

supported models: mtreg_l21,mtreg_lasso,mtreg_caso,mtreg_rmtl,mtreg_wmtl

supported datasets: computerBuyers,04cars,school,trafficSP,parkinson,isolet,sarcos,solarFlare

DATASET=syntheticWMTLR1 
MODEL=mtreg_l21 
ITERATIONS=1
GRID_SIZE=grid_small
MIN_TRAIN_SIZE=0.5
MAX_TRAIN_SIZE=0.5

./scripts/experiments.sh ${DATASET} ${MODEL} ${ITERATIONS} ${GRID_SIZE} ${MIN_TRAIN_SIZE} ${MAX_TRAIN_SIZE}

classification task experiments

supported models: mtclf_lasso,mtclf_l21,mtclf_caso,mtclf_srmtl,mtclf_wmtl

supported datasets: unrestnlp,crisisnlp,webKB,20news

DATASET=unrestnlp 
MODEL=mtclf_lasso 
ITERATIONS=10
GRID_SIZE=grid_small
MIN_TRAIN_SIZE=0.2
MAX_TRAIN_SIZE=0.9

./scripts/experiments.sh ${DATASET} ${MODEL} ${ITERATIONS} ${GRID_SIZE} ${MIN_TRAIN_SIZE} ${MAX_TRAIN_SIZE}

syntheticWMTLR4t20m100d25p0s0 syntheticWMTLR7t20m500d25p0s0 syntheticWMTLR8t100m100d1000p0s0

regression DATASETS=syntheticWMTLR4t20m100d25p0s0,school,04cars,computerBuyers,facebook,trafficSP,isolet MODELS=mtreg_lrst,mtreg_lasso,mtreg_l21,mtreg_caso,mtreg_rmtl,mtreg_wmtl,mtreg_swmtl3,mtreg_swmtl31,mtreg_swmtl32

classification DATASETS=syntheticWMTLC8t5m100d2p1s0,unrestnlp,crisisnlp,civilColombia,civilBrazil,civilMexico,civilParaguay,civilVenezuela MODELS=mtclf_lasso,mtclf_l21,mtclf_caso,mtclf_srmtl,mtclf_swmtl32,mtclf_swmtl33

''' MODELS=mtreg_swmtl33 DATASET=syntheticWMTLR8t20m100d25p1s0sp0 MODE=local STORE=local SOLVER=grid_small ITERS=5 MIN_TRAIN_SIZE=0.6 MAX_TRAIN_SIZE=0.6 ./scripts/run.sh $DATASET $MODELS $MODE $STORE $SOLVER $ITERS $MIN_TRAIN_SIZE $MAX_TRAIN_SIZE '''

''' MODELS=mtclf_caso DATASET=unrestParaguay TEST_DATASET=unrestParaguay2014 MODE=local STORE=local SOLVER=grid_small ITERS=1 MIN_TRAIN_SIZE=0.8 MAX_TRAIN_SIZE=0.8 ./scripts/run.sh $DATASET $MODELS $MODE $STORE $SOLVER $ITERS $MIN_TRAIN_SIZE $MAX_TRAIN_SIZE $TEST_DATASET '''

Reference and Citation

If you find this code useful in your research, please consider citing:

  @inproceedings{bai2023sign,
    title={Sign-Regularized Multi-Task Learning},
    author={Bai, Guangji and Torres, Johnny and Wang, Junxiang and Zhao, Liang and Abad, Cristina and Vaca, Carmen},
    booktitle={Proceedings of the 2023 SIAM International Conference on Data Mining (SDM)},
    pages={793--801},
    year={2023},
    organization={SIAM}
  }

BaiTheBest / SRML