Ordinal neural networks without iterative tuning

Area of Science:

• Machine Learning
• Artificial Intelligence
• Supervised Learning

Background:

• Ordinal regression (OR) is a crucial supervised learning technique bridging multiclass classification and regression.
• Traditional neural network models require iterative tuning, which can be computationally intensive.

Purpose of the Study:

• To adapt traditional neural network classification schemes for learning ordinal ranks.
• To develop an efficient OR model by imposing monotonicity constraints on neural network weights.

Main Methods:

• The proposed model incorporates monotonicity constraints by transcribing weights using padding variables, reformulating the problem as inequality constrained least squares (ICLS).
• The ICLS problem is solved analytically using a closed-form solution derived from Karush-Kuhn-Tucker conditions.
• Leveraging the extreme learning machine framework, input-to-hidden layer weights are randomly generated, eliminating iterative tuning.

Main Results:

• The model achieves competitive performance compared to existing state-of-the-art neural network methods for ordinal regression.
• The analytical solution provides an efficient parameter estimation without iterative tuning.

Conclusions:

• The proposed neural network model effectively handles ordinal regression tasks by enforcing monotonicity constraints.
• This approach offers an efficient and competitive alternative to traditional methods in ordinal rank learning.