Advances in Feature Selection with Mutual Information

摘要: The selection of features that are relevant for a prediction or classification problem is an important in many domains involving high-dimensional data. Selecting helps fighting the curse dimensionality, improving performances methods, and interpreting application. In nonlinear context, mutual information widely used as relevance criterion sets features. Nevertheless, it suffers from at least three major limitations: estimators depend on smoothing parameters, there no theoretically justified stopping feature greedy procedure, estimation itself dimensionality. This chapter shows how to deal with these problems. two first ones addressed by using resampling techniques provide statistical basis select estimator parameters stop search procedure. third one modifying into measure complementary (and not only informative) hand.