Coupled dimensionality reduction and classification for supervised and semi-supervised multilabel learning.

Area of Science:

• Machine Learning
• Data Science
• Statistical Modeling

Background:

• Coupled training of dimensionality reduction and classification improves single-label prediction.
• Multilabel learning presents unique challenges in prediction and dimensionality management.

Purpose of the Study:

• To develop a novel Bayesian method for supervised multilabel learning.
• To integrate linear dimensionality reduction with linear binary classification.
• To extend the method for intrinsic dimensionality estimation and semi-supervised learning.

Main Methods:

• A novel Bayesian method combining linear dimensionality reduction and linear binary classification for multilabel learning.
• A deterministic variational approximation algorithm for learning the probabilistic model.
• Extensions using automatic relevance determination for intrinsic dimensionality and a low-density assumption for semi-supervised learning.

Main Results:

• The proposed method achieves strong performance on benchmark multilabel datasets, measured by Hamming loss, average AUC, macro F1, and micro F1.
• Low-dimensional embeddings generated by the method are effective for exploratory data analysis.
• Demonstrated effectiveness in intrinsic subspace dimensionality determination and semi-supervised learning tasks.

Conclusions:

• The novel Bayesian approach offers a robust framework for multilabel learning, outperforming baseline methods.
• The method provides valuable insights through dimensionality reduction and supports semi-supervised learning scenarios.
• This work advances multilabel learning by integrating dimensionality reduction and classification within a unified probabilistic model.