Abstract: Due to the maldistribution of classes and features,the classification performance of traditional information gain algorithm will decline sharply．Considering that,a text feature selection method UDsIG was proposed which is based on the information gain．Firstly,because the feature selection may be influenced when the classes is unevenly distributed,we selected features based on class．Secondly,we used feature distribution uniformity to improve the influence on feature selection process when features are uneven distributed in the class．Then we adopt the feature relation tree model to deal with the class features,retain strong correlation features and delete the weak correlation and irrelevant ones．At last,we got the best feature subset by using of information gain formula which is based on weighted dispersion．The comparison experiment shows that the method has better classification performance.

Key words: Feature selection,Feature relation tree,Information gain,Imbalanced dataset,Dispersion