Determining Threshold Value on Information Gain Feature Selection to Increase Speed and Prediction Accuracy of Random Forest

Abstract Feature selection is a preprocessing technique aims to remove the unnecessary features and speed up the algorithm's work process. One of the feature selection techniques is by calculating the information gain value of each feature in a dataset. From the information gain value obtained, then the determined threshold value will be used to make feature selection. Generally, the threshold value is used freely, or using a value of 0.05. This study proposed the determination of the threshold value using the standard deviation of the information gain value generated by each feature in the dataset. The determination of this threshold value was tested on ten original datasets and datasets that had been transformed by FFT and IFFT, then classified using Random Forest. The results of the average value of accuracy and the average time required from the Random Forest classification using the proposed threshold value are better compared to the results of feature selection with a threshold value of 0.05 and the Correlation-Base Feature Selection algorithm. Likewise, the result of the average accuracy value of the proposed threshold using a transformed dataset in terms are better than the threshold value of 0.05 and the Correlation-Base Feature Selection algorithm. However, the calculation results for the average time required are higher (slower).