Partial Labelling Approached with through Structured Prediction (PLASP)

Topic

Generic implementations of weakly supervision algorithms developped in [CAB20], [CAB21a], [CAB21b].

Author

Vivien Cabannes

Version

1.0.0 of 2021/06/07

WARNING: For a fast implementation of (continuous) Laplacian spectral embedding, see https://github.com/VivienCabannes/laplacian

Installation

From wheel

You can download our package from its pypi repository.

$ pip install plasp

From source

You can download source code at https://github.com/VivienCabannes/partial_labelling/archive/master.zip. Once download, our packages can be install through the following command.

$ python <path to code folder>/setup.py install

You can also install it in develop mode, eventually with pip

$ cd <path to code folder>
$ pip install -e .

Usage

See files:

• problems/classification/libsvm_experiments.py
• problems/classification/semi_supervision_experiments.py
• and more generally *_experiements.py

Package Requirements

Most of the code is based on the following python libraries:

• numpy
• numba
• matplotlib

Some testing done with notebook are based on:

• jupyter-notebook
• ipywidgets

For ranking, we used the following lp solver library:

• cplex

To load LIBSVM files, more precisely to read libsvm files format we used:

• scikit-learn

To load MULAN files, more precisely to read mulan files format we used:

• arff
• skmultilearn

Datasets links

Datasets can be download at:

• LIBSVM: https://www.csie.ntu.edu.tw/~cjlin/libsvm/oldfiles/index-1.0.html
• MULAN: http://mulan.sourceforge.net/datasets-mlc.html

Change path in config file dataloader/config.py to specify path to your data.

See Also

A standalone package for fast computation of the Laplacian decomposition can be found at: https://github.com/VivienCabannes/laplacian

References

CAB20

Structured Prediction with Partial Labelling through the Infimum Loss, Cabannes et al., ICML, 2020

CAB21a

Disambiguation of weak supervision with exponential convergence rates, Cabannes et al., ICML, 2021

CAB21b

Overcoming the curse of dimensionality with Laplacian regularization in semi-supervised learning, Cabannes et al., NeurIPS, 2021