Semi-Supervised Learning Under General Causal Models.

Area of Science:

• Machine Learning
• Causal Inference
• Artificial Intelligence

Background:

• Semi-supervised learning (SSL) utilizes both labeled and unlabeled data for model training.
• The precise mechanisms by which unlabeled data enhance prediction accuracy remain incompletely understood.
• Causal perspectives offer a promising avenue for elucidating the role of unlabeled data in SSL.

Purpose of the Study:

• To propose a novel SSL framework capable of handling general causal models with flexible variable relationships.
• To investigate causal graph structures and develop corresponding causal generative models.
• To enhance predictive model accuracy by generating synthetic labeled data.

Main Methods:

• Developed an SSL framework accommodating complex, flexible causal relations between variables.
• Explored causal graph structures to inform the design of causal generative models.
• Utilized unlabeled data to learn these causal generative models.
• Generated synthetic labeled data from learned models for downstream prediction tasks.

Main Results:

• The proposed SSL framework effectively learns from unlabeled data under general causal assumptions.
• Learned causal generative models successfully produced synthetic labeled data.
• Empirical studies on simulated and real datasets demonstrated the effectiveness of the method in improving predictive accuracy.

Conclusions:

• Unlabeled data can significantly improve machine learning model performance within a causal framework, even with complex causal structures.
• The proposed method provides a robust approach to SSL by leveraging causal generative models.
• This work offers a new understanding and practical tool for utilizing unlabeled data in machine learning.