Personal view Semi-supervised learning is basically unsupervised learning with weights, so the "label" or "category" can be quantified rather than approximately clustered as occurs in unsupervised learning.
Unsupervised learning is PCA and tSNE and other multi-variate approaches. The advantages of semi-supervised learning is unlike unsupervised learning the feature in the data that shows greatest proportionate variance can be precisely identified and all features ranked; thus there aren't loads of components where the result could be anywhere.
Sure I've interpreted semi-supervised learning and found it was not useful for any comparative purposes. Basically the same variance Eigen-vector weight observed in one gene in one unique environment, is the strongest Eignen-vector weight for the same gene in a completely different environment. However, these are two independent semi-supervised learning analyses. Thus, it's not measuring comparative difference but relative variance between all genes in that given experiment.
Thus I personally didn't find it useful for my metric, HOWEVER such a result is good for some people where the overall structure of the variance identified by Eigen vectors is very similar between contrasting environments. Basically it's a comparison to all other genes in the analysis. To get it working you'd need to combine all genes in all environments into a single data set.
That would work, but ensure that the label for each feature (e.g. gene expression) comprised the environment and the gene. What the statistical theoretical validity of such a result would be - I haven't a clue - but it should give some interesting results. To be honest, if I'm not sure about the theoretical validity, I will not be alone.