Structure-Exploiting Discriminative Ordinal Multioutput Regression.

Area of Science:

• Machine Learning
• Statistics
• Data Mining

Background:

• Least-squares regression (LSR) has limitations in preserving class margins, impacting generalization on real-world data.
• Existing LSR variants that remodel regression targets show limited improvement due to unexploited structure-related information.

Purpose of the Study:

• To develop a novel multioutput regression model that exploits data structure for enhanced discrimination and generalization.
• To address the limitations of traditional regression techniques in handling structured data with ordinal labels.

Main Methods:

• Constructed a multioutput regression model utilizing a structure matrix to exploit intraclass correlations and input-output relationships.
• Incorporated a data-guided metric to discriminatively enlarge regression margins.
• Encoded model output as cumulative attributes for structured data with ordinal labels, forming the SEDOMOR model.
• Extended SEDOMOR to nonlinear versions using kernel functions and deep architectures, with derived optimization algorithms and convergence proofs.

Main Results:

• The proposed Structure-Exploiting Discriminative Ordinal Multioutput Regression (SEDOMOR) model effectively leverages data structure.
• SEDOMOR demonstrates superior performance and generalization ability compared to existing methods on extensive experiments.
• Nonlinear extensions of SEDOMOR further enhance distinguishing capabilities.

Conclusions:

• The developed SEDOMOR model and its extensions offer a powerful approach for discriminative ordinal multioutput regression.
• Exploiting data structure and enlarging regression margins are key to improving generalization in regression tasks.
• The proposed methods are effective and superior for handling structured data with ordinal labels.