Contrastive Learning for Semi-Supervised Deep Regression With Generalized Ordinal Rankings From Spectral Seriation.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Contrastive learning methods in regression rely heavily on labeled data.
• This dependency limits their application in semi-supervised settings where annotations are scarce.

Purpose of the Study:

• To extend contrastive regression methods for effective utilization of unlabeled data in semi-supervised learning.
• To reduce the dependence on costly data annotations for improved regression models.

Main Methods:

• Constructing a feature similarity matrix using both labeled and unlabeled samples within a mini-batch.
• Employing spectral seriation algorithms to recover ordinal relationships of unlabeled samples.
• Utilizing dynamic programming for robust feature selection to minimize perturbations.

Main Results:

• Demonstrated superior performance compared to state-of-the-art semi-supervised deep regression methods.
• Showcased the effectiveness of using recovered ordinal relationships for contrastive learning on unlabeled data.
• Validated theoretical guarantees and empirical results across various datasets.

Conclusions:

• The proposed method significantly enhances feature representation learning by incorporating unlabeled data.
• It offers a more reliable and robust approach to semi-supervised regression.
• The method provides an additional training signal for unlabeled samples, improving overall model accuracy.