Choosing pseudo-labels in co-training

Pick confident pseudo-labels from two views

Co-training (Blum and Mitchell, 1998) trains two classifiers on two feature views of the same examples. Each classifier labels the unlabeled points it is confident about, and those labels become training data for the other. This question covers only the selection step.

Implement pick_pseudo_labels(conf_a, pred_a, conf_b, pred_b, threshold).

• The four inputs are 1-D arrays of equal length N. conf_* holds each view's confidence in [0, 1], and pred_* holds its predicted class.
• A point is selected when at least one view has confidence at least threshold.
• If only one view is confident, use its prediction.
• If both views are confident and agree, use that label.
• If both views are confident and disagree, skip the point.
• Return (indices, labels): the selected indices in increasing order, and their labels.

Constraints

• All four arrays have the same length. Otherwise raise ValueError.
• threshold must be in (0, 1]. Otherwise raise ValueError.
• Confidences must lie in [0, 1]. Otherwise raise ValueError.
• Do not modify the inputs.

Hints