Incremental learning algorithm for large-scale semi-supervised ordinal regression.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Ordinal regression is a method for multi-class problems with ranked samples.
• Semi-supervised ordinal regression (SSOR) leverages unlabeled data for improved model training.
• Training large-scale SSOR models is challenging due to complex formulations and non-convexity.

Purpose of the Study:

• To develop an efficient algorithm for training large-scale semi-supervised ordinal regression models.
• To address the computational challenges associated with non-convex SSOR formulations.
• To provide a theoretical guarantee of convergence for the proposed SSOR algorithm.

Main Methods:

• An incremental learning algorithm for SSOR (IL-SSOR) is proposed.
• The algorithm updates SSOR solutions directly using KKT conditions.
• Finite convergence analysis is performed using the concave-convex procedure framework.

Main Results:

• IL-SSOR is the first efficient on-line learning algorithm for SSOR with guaranteed local minimum convergence.
• Experimental results demonstrate IL-SSOR achieves superior generalization compared to other semi-supervised multi-class algorithms.
• IL-SSOR shows similar generalization performance to other SSOR algorithms but with reduced computational time.

Conclusions:

• The proposed IL-SSOR algorithm effectively addresses the challenges of large-scale semi-supervised ordinal regression.
• IL-SSOR offers an efficient and convergent solution for training complex ordinal regression models.
• The algorithm provides a practical advancement for data mining applications requiring ranked classification.