Abduction in Machine Learning

Abduction and induction are strictly related forms of defeasible reason- ing. However, Machine Learning research is mainly focused on inductive techniques,leading from specific examples to general rules, with applica- tions to classification, diagnosis and program synthesis. Abduction has been used in Machine Learning, but its use was typically an aside tech- nique, to be integrated or added on top of the basic inductive scheme. We discuss the general relation between abductive and inductive reason- ing, showing that they solve different instantiations of the same problem. Then we analyze the specific ways abduction has been used in Machine Learning. Two different cases are individuated:(1) abductive reasoning used in explanation-based learning systems as a heuristic to guide search in top-down specialization, and (2) abduction used fo generating missing examples in relational learning. In both cases,the use of abduction is not general and adapted to a very tiny and specific problem. In this sense, the Machine Learning community has not used abduction as a synonym of induction, despite the high degree of similarity. However, both uses of abduction in learning have been proved to be effective for their intended purposes.