Feature Selection for Highly Skewed Sentiment Analysis Tasks

Sentiment analysis generally uses large feature sets based on a bag-of-words approach, which results in a situation where individual features are not very informative. In addition, many data sets tend to be heavily skewed. We approach this combination of challenges by investigating feature selection in order to reduce the large number of features to those that are discriminative. We examine the performance of five feature selection methods on two sentiment analysis data sets from different domains, each with different ratios of class imbalance. Our finding shows that feature selection is capable of improving the classification accuracy only in balanced or slightly skewed situations. However, it is difficult to mitigate high skewing ratios. We also conclude that there does not exist a single method that performs best across data sets and skewing ratios. However we found that TF ∗ IDF2 can help in identifying the minority class even in highly imbalanced cases.