Applying Relief Algorithm for Feature Selection in Sentiment Classification for Movie Reviews

摘要

This paper mainly focuses on using the Relief technique to improve the performance of sentiment classification (specifically the Na?ve Bayes classifier). We also present a way to find a suitable feature selection technique that can support the Na?ve Bayes classifier. We willcompare the Relief Algorithm with other feature selection methods that are currently considered innovative and which are featured in the latest research papers, namely the Information-Gain and Chi-Square techniques. In addition, we will present the calculation processes of the feature weightsof the data in order to compare the effectiveness of each feature selection method. The most commonly accepted approach to the sentiment classification problem is to first start by extracting and then selecting text features. Then the step in the feature selection process is to pick out thekeyword which is then to be used in the main preprocessing part that is critical to the indexing of documents. From the results of our study, it can be seen that utilizing the Na?ve Bayes Classifier without feature selection results in a relatively lower score. For the large movie reviewdataset, without feature selection we get a result of 78.80%. However, when the relief method is applied, the score is improved to 80.30%. In fact, when using the Na?ve Bayes Classifier with the Chi-Square and Information Gain methods separately, the score returned is slightly higheras well, giving off a score of 80.40%. In the future, improvements to these methods could be made but along with the improvements in our technology, we may witness more efficient methods to sentiment analysis due to either more efficient algorithms or a more rapid execution of algorithms dueto improved computational executions. Eventually, the goal would be to surpass the 80% region and to eventually break into at least the 90% region of accuracy.