Cauchy-MMA: Multi-Metric autoencoder ensemble with adaptive cauchy loss for High-Dimensional incomplete data...

Area of Science:

• Machine Learning
• Data Science
• Recommender Systems

Background:

• High-dimensional incomplete data presents challenges in collaborative filtering due to outliers and static loss functions.
• Existing multi-metric autoencoders struggle to dynamically balance robustness and generalization.

Purpose of the Study:

• To propose Cauchy-M MA, a novel framework enhancing collaborative filtering performance.
• To address limitations of static loss functions in multi-metric autoencoders for robust representation learning.

Main Methods:

• Integrated multi-metric autoencoders with adaptive Cauchy loss for outlier resistance.
• Developed a continuous, differentiable parameter adaptation mechanism for dynamic penalty adjustment.
• Created a six-branch collaborative learning system by combining adaptive Cauchy loss variants with static losses.

Main Results:

• Theoretical analysis confirmed sublinear regret bounds and asymptotic convergence.
• Cauchy-MMA demonstrated competitive performance against graph-based and robust baselines.
• The framework significantly mitigated performance degradation caused by strong outlier interference.

Conclusions:

• Cauchy-MMA offers a robust and adaptive solution for collaborative filtering with high-dimensional incomplete data.
• The adaptive Cauchy loss effectively suppresses outliers and improves generalization.
• The proposed method achieves superior performance in rating prediction and Top-K recommendation tasks.