# Predicting probability for each tag given already chosen tags

I have a set of tags (~10'000, will be extended over time) presented to a user. After he has selected 3 or more tags, I want to predict for each remaining tag what the chances are that the user will select this tag as well. I strictly need the prediction for all remaining tags. The prediction should be fast but accuracy isn't that important. The training data would be cases where users already have selected a subset of tags. What would be a good approach here?