Ranking and Combining Latent Structured Predictive Scores without Labeled Data.

Area of Science:

• Machine Learning
• Bioinformatics
• Data Science

Background:

• Combining predictors from distributed data sources enhances prediction accuracy.
• Assessing predictor accuracy typically requires extensive labeled data, which is often difficult to acquire.
• Correlated predictors, common in ensemble learning, pose integration challenges.

Purpose of the Study:

• To develop a novel structured unsupervised ensemble learning (SUEL) model for integrating predictors without labeled data.
• To address the challenge of unknown predictor accuracy and high predictor correlation.
• To rank and combine predictors effectively for improved meta-learner performance.

Main Methods:

• Introduced a novel structured unsupervised ensemble learning (SUEL) model.
• Developed two correlation-based decomposition algorithms: constrained quadratic optimization (SUEL.CQO) and matrix-factorization-based (SUEL.MF).
• Evaluated the SUEL model using simulation studies and a real-world risk gene discovery application.

Main Results:

• The SUEL model successfully ranks predictors without requiring ground truth data.
• The proposed SUEL.CQO and SUEL.MF methods efficiently estimate the SUEL model.
• The ensemble model integrates dependent predictors effectively, demonstrating enhanced performance.

Conclusions:

• The proposed SUEL methods provide an effective solution for integrating correlated predictors without labeled data.
• This approach enhances meta-learner performance in prediction problems with limited ground truth.
• The methods show promise for applications like risk gene discovery in bioinformatics.