Repository for reproducing the results of the paper "Dependence and Model Selection in LLP: The Problem of Variants" from KDD23

The Split-Bag methods proposed at this paper are implemented at llp-learn

• Python 3.8 or higher (developed on Python 3.8)
• R version 4.2.1

pip3 install -r requirements.txt

To use LMM¹ it is necessary to get its code:

git clone https://github.com/giorgiop/almostnolabel.git

To install the R libraries:

install_r_libraries.py

python3 kdd_experiment.py -d {dataset_name} -m {model} -l {loss} -n {n_splits} -v {validation_size_percentage} -s {splitter} -e {execution_number}

As an example, we have:

python3 kdd_experiment.py -d mnist-digits-6-7-naive-3bags-cluster-None-None -m lmm -l abs -n 3 -v 0.5 -s split-bag-bootstrap -e 0

For $k$-fold based methods, the validation_size_percentage is not used

python3 kdd_experiment.py -d mnist-digits-6-7-naive-3bags-cluster-None-None -m lmm -l abs -n 3 -s split-bag-k-fold -e 0

./paper_run_all_experiments.sh

Each execution produces one parquet file. After running all the experiments, they can be combined into one single file (kdd-experiment-results.parquet) as following:

python3 aggregate_results.py

./paper_plot_results.sh

The plots are saved in the plots folder.

./paper_table_results.sh

The tables are saved in the tables folder.