Sentiment Prediction Accuracy of Amazon Fine Food Review using TF-IDF and LightGBM models

Abstract

Changes in the pattern of society in meeting their needs develop as the times progress from conventional to digital. This makes service providers need to change business work patterns towards digitizing buying and selling transactions. Service providers serve consumer needs digitally and maintain optimal service patterns. One of the efforts to maintain optimal service is through community response to services, both positive and negative. The community response can be analyzed using sentiment analysis. This study focuses on the analysis of the accuracy of sentiment predictions on the Amazon fine food review dataset, which was taken as many as 20,000 data samples. The analysis was carried out in various stages, namely dataset collection, data preprocessing, TF-IDF, and LightGBM. The test results used TF-IDF and LightGBM with TF-IDF parameter settings of 1 to 2 grams and LightGBM parameter settings with a max_depth of 50. Num_leaves used were 40 and the learning rate was 0.1 on the Amazon Review dataset which took 20,000 samples. The analysis carried out resulted in a predictive level of sentiment accuracy above 90%, reaching 93.2%.