PU.py

from sklearn.svm import SVC
from sklearn.metrics import accuracy_score
import torch
import numpy as np
    
def pu_classification(embeddings_positive, embeddings_unlabeled, iterations=10, selection_ratio=0.3):
    y_train_pu = np.ones(embeddings_positive.shape[0])  # Positive examples
    y_unlabeled_pu = -np.ones(embeddings_unlabeled.shape[0])  # Unlabeled examples

        # Convert numpy arrays to PyTorch tensors
    embeddings_positive = torch.tensor(embeddings_positive)
    embeddings_unlabeled = torch.tensor(embeddings_unlabeled)

    X_train_pu = torch.cat((embeddings_positive, embeddings_unlabeled), dim=0)
    y_train_pu = np.hstack((y_train_pu, y_unlabeled_pu))

    pu_classifier = SVC(kernel='linear', probability=True)

    for i in range(iterations):
        pu_classifier.fit(X_train_pu.numpy(), y_train_pu)

            # Select bottom samples from unlabeled data based on confidence and add them to positive samples
        unlabeled_probs = pu_classifier.predict_proba(embeddings_unlabeled.numpy())[:, 1]
        bottom_indices = unlabeled_probs.argsort()[:int(selection_ratio * embeddings_unlabeled.shape[0])]
        selected_samples = embeddings_unlabeled[bottom_indices]

        embeddings_positive = torch.cat((embeddings_positive, selected_samples), dim=0)
        embeddings_unlabeled = np.delete(embeddings_unlabeled, bottom_indices, axis=0)

        X_train_pu = torch.cat((embeddings_positive, embeddings_unlabeled), dim=0)
        y_train_pu = np.hstack((np.ones(embeddings_positive.shape[0]), -np.ones(embeddings_unlabeled.shape[0])))

    return pu_classifier