Bayesian Optimization for Parameter Selection of Random Forests Based Text Classifier

5 While random forest algorithm has been found to be prominent for various 6 classification tasks, like many other machine learning algorithms it requires 7 a number of parameters to be tuned to ensure better performance. Even 8 though the strong influence of different parameters on random forest is 9 evident, an attempt to systematically optimize these parameters is rare. 10 Common techniques for parameter tuning such as cross validations are not 11 often sufficient in this case, as the number of choices are increased. In this 12 context, we propose a Bayesian optimization method to tune the parameters 13 of random forest. We implemented a text classification system using the 14 random forest package of Scikit-learn. To evaluate our approach, we 15 compare the results on different parameter settings generated during 16 optimization procedure. We also examine how various choices of 17 acquisition functions could potentially affect the optimization. Our results 18 suggest that by tuning the parameters for random forest, we could enhance 19 the classification performance over default choices of parameters provided 20 in Scikit-learn package. 21