Learning Output Relevant Features by Joint Autoencoder.

Area of Science:

• Machine Learning
• Data Science
• Statistical Modeling

Background:

• Classical regression models can be hindered by irrelevant input information, impacting predictive performance.
• Identifying and utilizing output-relevant features is crucial for enhancing regression accuracy.

Purpose of the Study:

• To propose a novel joint autoencoder (JAE) for learning output-relevant features in regression tasks.
• To improve predictive and reconstructive performance by integrating supervised and unsupervised learning mechanisms.
• To enable hierarchical feature representation learning for enhanced model generalization.

Main Methods:

• Developed a joint autoencoder (JAE) that combines supervised and unsupervised learning.
• Implemented a local parameter embedding strategy for successive JAEs to preserve predictive performance in hierarchical structures.
• Evaluated the model's performance on nine diverse datasets with varying sample sizes.

Main Results:

• The proposed JAE effectively learns output-relevant features, improving predictive performance.
• The model demonstrated enhanced reconstructive capabilities, leading to better generalization on unseen data.
• Experimental results confirmed the predictive performance and robustness of the JAE across multiple datasets.

Conclusions:

• The joint autoencoder (JAE) offers a robust approach to enhance regression models by focusing on output-relevant features.
• Hierarchical representations learned through successive JAEs maintain predictive accuracy.
• The method provides a promising direction for improving the efficiency and effectiveness of regression analysis.