Does imputing model labels using the model predictions can improve it’s performance?

In some scenarios a data scientist may want to train a model for which there exists an abundance of observations, but only a small fraction of is labeled, making the sample size available to train the model rather small. Although there’s plenty of literature on the subject (e.g. “Active learning”, “Semi-supervised learning” etc) one may be tempted (maybe due to fast approaching deadlines) to train a model with the labelled data and use it to impute the missing labels. While for some the above suggestion might seem simply incorrect, I have encountered such suggestions on several occasions and had a hard time refuting them. To make sure it wasn’t just the type of places I work at I went and asked around in 2 Israeli (sorry non Hebrew readers) machine learning oriented Facebook groups about their opinion: Machine & Deep learning Israel and Statistics and probability group. While many were referring me to methods discussed in the literature, almost no one indicated the proposed method was utterly wrong. I decided to perform a simulation study to get a definitive answer once and for all. If you’re interested in reading what were the results see my analysis on Github.