I have a data collection with a mixed feature set consisting of both numerical features and text features. The number of numerical features is quite small, i.e., 6, comparing to the number of text features, i.e., 2000. ( The data set is a csv file, where there are 6 columns consisting of digit values only; other columns are just plain texts). We found that removing these numerical features (columns) in fact improves the performance. However, when we conduct the feature selection based on the information gain, looks like those numerical features gets more weight. How to explain this scenario?