Multiple Ordinal Regression by Maximizing the Sum of Margins

Area of Science:

• Machine Learning
• Computational Psychology
• Data Science

Background:

• Human preferences are typically measured using ordinal variables, requiring systems to learn orderings from data.
• Existing ordinal regression methods using support vector machines (SVMs) face computational challenges or limitations in handling multiple ranking variables.
• Current approaches often rely on a single ranking variable, restricting the complexity of learned preferences.

Purpose of the Study:

• To develop a computationally efficient and flexible multiple ordinal regression algorithm for estimating human preferences.
• To address the limitations of existing methods by incorporating multiple ranking variables into the SVM framework.
• To apply the novel algorithm to model ordinal associations between attributes and object categories in human subjects.

Main Methods:

• Developed a multiple ordinal regression algorithm by maximizing the sum of margins between consecutive classes across multiple rankings.
• Derived an efficient, iterative solution using a sequential minimal optimization (SMO) procedure for practical implementation.
• Utilized support vector machine (SVM) principles with common direction vectors and distinct biases for class separation.

Main Results:

• The proposed algorithm provides an efficient and easy-to-implement solution for ordinal regression problems.
• Demonstrated the accuracy of the algorithm across several diverse datasets.
• Showcased superior performance of the developed ordinal associations compared to traditional binary methods in attribute association tasks.

Conclusions:

• The novel multiple ordinal regression algorithm effectively estimates human preferences and decision mechanisms using multiple rankings.
• The sequential minimal optimization approach offers a computationally efficient alternative to existing, more complex methods.
• The application in human attribute-object association tasks highlights the practical utility and improved accuracy of the proposed method over binary approaches.