As @shadowtalker indicated, a binary features indicating the existence of co-borrower should be helpful anyway. As you certainly know, "imputing missing values" as a procedure for guessing value of missing values is absolutely meaningless as it produces values for something which do not exist. If it's about salary and similar numeric data, just simply put 0 (or the neutral value. Whatever it is.) and if it's about nominal variables, add a new value equivalent to None. at the end, all single customers must have exactly same values for these features.
Also one practical idea could be dividing your data to single borrowers and 2-borrowers at the very beginning of classification so the whole model will become like this:
Above model is somehow equivalent to the feature-selection property of some classifiers such that Decision Trees. In a Bayesian Network, such a feature could be the root node.