The authors tested their models using the Amazon Mechanical Turk (AMT) synthetic fake reviews dataset on a real-world fake reviews dataset procured from Yelp.
The addition of behavioral features increases their accuracy to 86.1 % on Yelp's filtered reviews dataset.
For example, when using the same framework to evaluate the artificial AMT dataset used in [3, 12, 25] and Yelp's filtered reviews dataset, the extracted features and results differed greatly, especially when using n-gram text features [23].
We use Turkish and English review datasets in our experiments.
DFD does reasonably well across all review datasets, but it tends to favour bigger feature sizes.
We have grouped collected review datasets into reviews by the recommended review group (non-spam) and reviews by the fake review group (filtered).
Also compared with the Turkish review datasets, DFD is not as good as CHI2 and IG for the English review datasets, even though the performance is close for the kitchen reviews and generally better than OCFS.
As can be seen in Table 1, the total number of features without any reduction ranges from 9000 to 18,000 for the Turkish review datasets, and 8,000 38,000 for the English review datasets.
Our results show that for all Turkish review datasets, the best results are all obtained with the NBM classifier, and for some English review datasets, LR and SVM have the best performance.
Feature sizes are selected in the range from 500 to 3000 with 500 increments, compared with the total feature sizes ranging from 8000 to 18,000 for the Turkish review datasets and from 8000 to 38,000 for English review datasets.
For differences, the English review datasets usually have bigger vocabulary, resulting in relatively bigger feature sizes for feature selection.
