Feature Selection Method Based on Hybrid SA-GA and Random Forests
Zibo Zhou, Yunfan Wang, Man Li · 2020
We propose a feature selection algorithm based on hybrid simulated annealing (SA) - genetic algorithm (GA) and random forests, whose procedure can be described as the following steps. First, set an initial temperature and create an initial solution as current solution by binary encoding of the features. Then, generate a new potential solution according to the principle of crossover and mutation in the genetic algorithm. Next, construct random forests and calculate the fitness degree of the new potential solution by using out-of-bag (OOB) error. The acceptance of new state is based on Metropolis criterion, and the steps are iterated until a predefined lowest temperature is reached. Finally, a contrast experiment has done to show that the hybrid algorithm we proposed has better classification and feature selection performance than the methods from the existing literature.