Confident Sinkhorn Allocation for Pseudo-Labeling

Overview | Installations | Reproduce Experiments | Make Plot | Public Datasets | References

✨ Overview

We propose a semi-supervised learning method for tabular data that does not require any domain-specific assumption. The method we propose is based on pseudo-labeling of a set of unlabeled data using Confident Sinkhorn Allocation (CSA). Our method is theoretically driven by the role of uncertainty in robust label assignment in SSL. CSA will assign labels to only the data samples with high confidence scores using Sinkhorn’s algorithm. By learning the label assignment with optimal transport, CSA eliminates the need to predefine the heuristic thresholds used in existing pseudo-labeling methods, which can be greedy. The proposed CSA is applicable to any data domain, and could be used in concert with consistency-based approaches, but is particularly useful for tabular data where pretext tasks and data augmentation are not applicable.

🔨 Installations

pip install -r requirements.txt

Packages

* colorama==0.4.5
* cycler==0.11.0
* fonttools==4.33.3
* joblib==1.1.0
* kiwisolver==1.4.3
* matplotlib==3.5.2
* numpy==1.23.0
* packaging==21.3
* pandas==1.4.3
* Pillow==9.2.0
* pyparsing==3.0.9
* python-dateutil==2.8.2
* pytz==2022.1
* scikit-learn==1.1.1
* scipy==1.8.1
* six==1.16.0
* threadpoolctl==3.1.0
* tqdm==4.64.0
* xgboost==1.6.1

➡️ Run and reproduce the experiments

• Reproduce the experiments

• Demo Confident Sinkhorn Allocation for Multiclassification

• Demo Confident Sinkhorn Allocation for Multilabel

Run single experiment. The result pickle file will be stored in --save_dir

• run CSA

cd run_experiments
python run_csa.py --numIters 5 --numTrials 5 --numXGBs 10 --confidence_choice "ttest" --dataset_name "synthetic_control_6c" --save_dir "../results_output"

• run Pseudo labeling

cd run_experiments
python run_pseudo_labeling.py --numIters 5 --numTrials 5 --upper_threshold 0.8 --dataset_name "synthetic_control_6c" --save_dir "../results_output"

• run FlexMatch

cd run_experiments
python run_flexmatch.py --numIters 5 --numTrials 5 --upper_threshold 0.8 --dataset_name "synthetic_control_6c" --save_dir "../results_output"

• run UPS

cd run_experiments
python run_ups.py --numIters 5 --numTrials 5 --numXGBs 10 --upper_threshold 0.8 --lower_threshold 0.2 --dataset_name "synthetic_control_6c" --save_dir "../results_output"

• run SLA

cd run_experiments
python run_sla.py --numIters 5 --numTrials 5 --numXGBs 10 --confidence_choice "None" --dataset_name "synthetic_control_6c" --save_dir "../results_output"

Plot results

Please specify the following parameters. These parameters will link to the correct files in your result folders

* save_dir = 'results_output' # path to the folder store the results 
* out_file='' # change this if you have set it when running the experiments 
* numTrials=20 # number of repeated trials
* numIters=5 # number of used pseudo-iterations
* dataset_name='madelon_no' # datasets
* list_algorithms=['Pseudo_Labeling','FlexMatch','UPS','CSA'] # list of algorithms to be plotted

the following parameters to be used to load the correct paths

* confidence='ttest' # only used for CSA 
* upper_threshold=0.8 # only used for Pseudo_Labeling,FlexMatch
* low_threshold=0.2 # only used for UPS
* num_XGB_models=10 # only used for CSA and UPS

python plot_results.py

Datasets for multiclassification from UCI

we can load all datasets for multiclassification with their datasetName as follows:

with open('all_data.pickle', 'rb') as handle:
    [all_data, datasetName_list] = pickle.load(handle)

List of all datasets

* segment_2310_20
* wdbc_569_31
* analcatdata_authorship
* synthetic_control_6c
* German-credit
* madelon_no
* agaricus-lepiota
* breast_cancer
* digits

Datasets for multilabel classification from UCI

we can load all datasets for multi-label classification with their datasetName as follows:

with open('all_data_multilabel.pickle', 'rb') as handle:
    [all_data, datasetName_list] = pickle.load(handle)

List of all datasets

* yeast
* emotions

✏️ References:

Vu Nguyen, Sachin Farfade, and Anton van den Hengel. "Confident Sinkhorn Allocation for Pseudo-Labeling." Workshop on Distribution-Free Uncertainty Quantification at ICML 2022