Transductive ordinal regression.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Ordinal regression is often treated as a multiclass problem, facing challenges with increasing class numbers due to the need for extensive labeled data.
• Labeled data for ordinal regression can be costly or difficult to acquire, while unlabeled data is typically abundant.

Purpose of the Study:

• To introduce a novel transductive learning paradigm for ordinal regression (TOR) that utilizes readily available unlabeled data.
• To address the challenge of simultaneously estimating unlabeled data's ordinal class labels and decision functions.

Main Methods:

• Developed a transductive ordinal regression (TOR) framework with an objective function adaptable to various loss functions in transductive settings.
• Introduced a label swapping scheme to ensure a strictly monotonic decrease in the objective function value.

Main Results:

• Demonstrated the effectiveness of TOR through extensive numerical studies on benchmark datasets, including sentiment prediction.
• Showcased the robust and improved performance of TOR compared to state-of-the-art ordinal regression techniques.

Conclusions:

• The proposed transductive learning paradigm for ordinal regression (TOR) offers a robust and effective approach to enhance performance.
• TOR successfully leverages unlabeled data to overcome limitations associated with labeled data scarcity in ordinal regression tasks.