Hyper label model

A hyper label model to aggregate weak labels from multiple weak supervision sources to infer the ground-truth labels in a single forward pass.

For more details, see our ICLR23 paper Learning Hyper Label Model for Programmatic Weak Supervision

** To reproduce experiments of our paper or to re-train the model from scratch, please switch to the paper_experiments branch.

How to use

1. Install the package

   pip install hyperlm

2. Import and create an instance

   from hyperlm import HyperLabelModel
   hlm = HyperLabelModel()

3. Unsupervised label aggregation. Given an weak label matrix X, e.g. X=[[0, 0, 1], [1, 1, 1], [-1, 1, 0], [0, 1, 0]], you can infer the labels by:

   pred = hlm.infer(X)

Note in X, -1 represents abstention, 0 and 1 represent classes. Each row of X includes the weak labels for a data point, and each column of X includes the weak labels from a labeling function (LF).

4. Semi-supervised label aggregation. Let's say the gt labels are provided for the examples at index 1 and 3, i.e. y_indices=[1,3], and the gt labels are y_vals=[1, 0]. We can incorporate the provided partial ground-truth with:

   pred = hlm.infer(X, y_indices=y_indices,y_vals=y_vals)

Citation

@inproceedings{
wu2023learning,
title={Learning Hyper Label Model for Programmatic Weak Supervision},
author={Renzhi Wu and Shen-En Chen and Jieyu Zhang and Xu Chu},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=aCQt_BrkSjC}
}