Semi Supervised Learning

Demonstrate how adding unlabeled data to a supervised problem can improve out-of-sample accuracy.

This repo provides an example of this accuracy improvement through a loss function := labeled_loss + unlabeled_loss which is optimized using expectation-maximization (EM).

Follows Nigam et al 2006 from "Semi-Supervised Learning", Chapelle et al.

Running experiments

Create virtual env

python -m venv venv4ssl

Activate virtual env

Windows/Anaconda: venv4ssl/Script/activate

Powershell notes:

you may need to Powershell as administrator
\venv4ssl\Scripts\Activate.ps1

macos/Linux: source venv4ssl/bin/activate

Linux notes: installing matplotlib may require headers like ft2build.h

Install requirements

python -m pip install --upgrade pip
pip install -r requirements.txt

nltk_data note:

Handled by specifying nltk data dir in run_experiments_main

run experiments with varying num of labeled samples:

python src/run_experiments_main.py 
--n_labeled 20,100,300,500,700,1000 
--n_unlabeled 10000 
--max_iters 5 
--out_dir <your_output_dir> 
--nltk_data_dir <optional: local nltk data dir> 
--test_acc_plot_fname test_acc.png