Guided Class Expression Learning

Class Expression Learning is a challenging but important task. Given sets of positive and negative examples, the goal is to generate a class expression that includes as many positive examples as possible but excludes the negative examples. However, the search for such a class expression happens in an infinite space and, hence, can become quite challenging. A classic approach to this problem is a loop in which the currently best known expression is chosen and further refined using a refinement operator. The image below shows such a tree of refinements. The green nodes are the ones that are chosen for a further refinement.

Although current approaches can work quite well, they tend to guide the search based on a summary of the expression's performances while the program itself would be able to provide more detailed information to guide the search. Hence, the goal of this thesis is to make the search more intelligent and choose expressions or their combinations in a data-driven way. At the same time, it might be necessary to introduce additional measures for the quality of expressions to avoid overfitting.

Prerequisites:

• Java programming skills can be helpful
• A basic understanding of Description Logics