Semi-Supervised Learning.pdf

Semi-supervised learning (SSL) combines labelled and unlabelled data to improve predictions. The success of SSL is linked to the principle of independent causal mechanisms, which suggests that SSL is possible when predicting a target variable from its effects, but not from its causes. This work extends the investigation of connections between SSL and causality to a more general setting, where both cause and effect features are used to predict a target variable.

The key insight is that the relevant information for prediction is contained in the conditional distribution of effect features given causal features. This generalizes previous results for causal and anticausal learning. The work reformulates classical SSL assumptions and proposes algorithms based on these assumptions.

The main contributions are:
1. Extension of the investigation of connections between SSL and causality to a more general setting.
2. Identification of the conditional distribution of effect features given causal features as the relevant information for prediction.
3. Reformulation of classical SSL assumptions and proposal of new algorithms.

The work is evaluated on synthetic and medical datasets, and the results empirically support the analysis. The assumptions and results are critically discussed, and an outlook on future work is provided.

The key concepts include:
1. Semi-supervised learning (SSL)
2. Causality and independent causal mechanisms
3. Conditional cluster assumption
4. Cause and effect features
5. Conditional distribution of effect features given causal features.

The main methods used are:
1. Self-learning (Yarowsky-algorithm)
2. Generative model approaches
3. Graph-based approaches
4. Transductive SVMs.

The key findings are:
1. SSL is possible when predicting a target variable from its effects, but not from its causes.
2. The conditional distribution of effect features given causal features contains the relevant information for prediction.
3. The proposed algorithms outperform well-established SSL algorithms on synthetic and medical datasets.