Extraction Features Using Lexicon Embedding On Convolutional Neural Network For Fake Review Detection

Opinion review became important thing for both individuals and organizations. Although opinions were subjective, the collection of opinions had views, experiences, and interests that made opinions facilitate individuals in making buying decisions and organizations in improving their productions from existing feedback. However, many entities took advantage of doing deception by creating fake reviews. In this research, the authors created a fake review detection system using the Convolutional Neural Network (CNN) method with feature extraction. The feature extractions included word embedding and lexicon embedding with naïve concatenation, multichannel, and separated convolution integration. The research aimed to obtain accuracy, precision, recall, and F1score from both features and with unbalanced and balanced classes. The authors used literature study to collect data and perform simulations. The authors’ models were experimented with both 6,632 hotel reviews and 60,763 restaurant reviews in the Yelp filtered datasets, with a ratio of 1 fake review to 6 genuine reviews in both topics. Our analysis of quantitative data revealed that the embedding lexicon feature had significant effect on base CNN model. The lexicon embedding separated convolution exhibited the best overall performance, achieving 78% accuracy, 81% recall, 74% precision, and 77% F1-score in balanced classes.