A Random Feature Selection Approach for Neural Network Ensembles: Considering Diversity
Junfei Che, Qingfeng Wu, Huailin Dong · 2010
The concept of ensemble feature selection has been raised by Optiz in his earlier work. And yet, for models like neural networks, new models should be trained and created for every change in its feature subspace, this problem may become tricky when evolutionary algorithms are used to select features, for the slow-training process of neural networks may dramatically extend the whole process of ensemble training. Given the success of a powerful ensemble approach - GASEN, a random feature selection method is adopted to solve this problem. Experiments show that this approach (GASEN-fs) not only accelerate the training of component networks but also enhance its generalization ability.