The comments contained on e-commerce users generally contain opinions about positive or negative experiences at several online shops. Sentences that can be written indirectly both a little or a lot, will affect other potential customers. So as a result of these comments cause a product sold at an online store has a rating of two things namely "recommended" or "non-recommended". However, detection of positive and negative opinions manually will require more time because of the large amount of data. For this reason opinion mining using technology in data mining can be used to automate positive and negative detection of comments. However, one of the main problems in opinion mining is limited data but has a large number of attributes. In this study, we propose the application of Pearson correlation (PC) based feature selection for opinion mining optimization. The results of the experiment show that the application of PC increases the performance of opinion mining systems in 3 types of classification, namely Logistic Regression, Naïve Bayes and Support Vector Machine, resulting in more optimal accuracy, namely 98.80%, 87.87% and 98.12%.