What Can and What Has To Be Learned From Data

Endre Boros

Not diminishing the importance of experimental/expectation based evaluation of learning approaches, we think that there is also a need for more exact evaluations. After all, not only there are some regularities which we may learn from data, there might be some special structures hidden in the data which we may view as necessary not to miss. For rule based learning approaches we propose a notion of “justifiability”, and try to argue that this notion is compatible with existing methods, does not lead to over-fitting, and helps to identify of what a “good and professional” learning method is.

(Joint (old) work with Yves Crama, Peter L. Hammer, Toshi Ibaraki, Alex Kogan, and Kaz Makino.)