Abstract:

Classification performance of a traditional information gain algorithm will rapidly decline when feature items are in an unbalanced distribution. An improved calculation method of an information gain formula using feature items’ distribution information is proposed. Distribution information of feature items is computed to judge whether the imbalance of feature items exists and balance the influence of classification accuracy when the feature items do not appear. The improved calculation method has better performance through the experiment.

Key words: feature selection; information gain; distribution Information inside a class; distribution Information among classes