A Relevance Relation is a predictor that maps two or more software entities to a value r quantifying how relevant i.e. connected or related, the entities are to each other. In other words r shows the strength of relevance among the entities. In this talk we present our research involving learning maintenance relevance relation between files in a large legacy system, where change in one file may require a change in another. We will discuss some of experiments we have performed and their results. We will also discuss results of learning in the presence of different misclassification costs, and present one way of comparing performance of models in terms of error costs. We hope to end this presentation with a brain storming session where audience can provide us with suggestions specially as to how best one can simulate/analyze the performance of a generated model in settings such as the ones discussed.