Aspect Based Sentiment Analysis in E-Commerce User Reviews Using Latent Dirichlet Allocation (LDA) and Sentiment Lexicon

User ratings on products sold by e-commerce greatly influence the number of purchases. Positive ratings will encourage other buyers to participate in buying the product. While negative ratings given by users will reduce the interest in purchasing products. Nonconformities between rating and user reviews sometimes provide a wrong assessment of a product. This happens because buyers also provide reviews on the quality of delivery services from e-commerce. Based on that issue, the utilization of the Latent Dirichlet Allocation (LDA) could be used on sentiment analysis of the user reviews. Sentiment analysis of the user reviews aims to facilitate e-commerce in informing the product quality as rating supporters that have been given by users. This research aims to determine the classification performance of sentiment analysis on e-commerce user reviews using the LDA algorithm with input data in the form of e-commerce user reviews. Then, compare the application of sentiment analysis of the user reviews with the use of general training data and per category training data. The result of this research showed that in the first iteration the best architecture was produced by the application of LDA with a combination of parameters of alpha 0.001, beta 0.001, and number of topics 15. The architecture had 67,5% accuracy level. From the best architecture then training data input is given based on each product review category. The result showed that the combination of the usage of general data and per category data indicate an increase in the average accuracy of 0,82% from the three-test data. Therefore, in order to produce the best performance of building a classification model of sentiment analysis of the user reviews, it should be performed by applying LDA with a combination of general data and per category data usage.