Yelp

Yelp fake review detection AI

Dataset provider

For the detection of fake/true reviews I have used the dataset provided by professor Rayana and Akoglu. This was a big win for me as I do not have to analyse and somewhat guess which reviews are fake and which are true. Yelp has their own way of finding fake reviews, but its closed source. It includes user, product info, ratings and the review itself. However I suppose I will only concentrate on the review features for now, as I do not have much time to focus on the rest. Fake reviews are labelled as -1 and true are labelled as 1. This dataset has around 300,000 reviews which is more that enough to get a reliable rating. The research paper that I am going to follow is titled “Fake Review Detection on Yelp”, done by students at Stanford. Although it does not have code for me to look at and learn, I suppose it would be fun to try and sort of implement it on my own.

Progress

I’ve only read the research paper for now, but I am expecting to complete this algorithm by this week. The way the research paper finds out which and fake/real is that they use neural networks. However there is one bump that I have run into, which is Linear Discriminant Analysis, which is something new. From what I understand, it is a way to reduce dimensions in a graph to its lowest possible form, while not losing much information. The use for this is for choosing uni and bigrams from the reviews. More research needs to be done, as I also do not know why bi and unigrams are used for in NLP exactly.Although there is one more way which is to just take the top 100 unigram and bigram words, although easier, it leads to a poorer performance.

Perhaps if the use of uni and bi grams is for the AI to learn the context of the word, then I suppose word embedding would be a better choice.

End

Ok that’s all for now, see you in the next blog post!