Fowarding the Choice of Bias Legal-f: Using Feature Selection to Reduce the Complexity of Legal

Preprocessing data is a means for improving the computational eeciency of concept learning systems. The selection of relevant features increasingly becomes a central problem in machine learning. Two diierent approaches were intensively studied these last few years, the lter and the wrapper methods. Both methods wish to give a learner only those attributes that are relevant to the target concept. Lattice based machine learning systems seems to better represent concepts inside the data. Several systems using this framework have been developed recently. Their principal drawback comes from the exponential complexity of the lattice construction. In this paper, we present a method to select relevant features according to the learning algorithm criteria. A rst application is made to bring up a new lattice-based machine learning system, LEGAL-F, which decreases considerably the computational complexity of the previous LEGAL system. A study of the theoretical complexity of LEGAL-F is done, and experimental comparisons are reported rst between LEGAL and LEGAL-F in terms of complexity, and then between LEGAL-F and C4.5 in terms of classiication accuracy.