Learn To Rate Fine Food

We investigate a food review dataset from Amazon with more than 500000 instances, and utilize information from the data set such as the text review, score, helpfulness, etc. Instead of the traditional word representation using frequency, we use skip-gram to train our own word vectors using the pretrained GloVe Twitter word vector as the initialized value. We also use recursive parsing tree to train the vector of the whole sentence with the input of word vectors. After that, we use neural network methods to classify our review text to different scores. We mainly research and compare the performance of Gated Recurrent Unit Network (GRU) and Convolutional Neural Network (CNN) on our data. After tuning the hyper parameters, we get our best classifier as a bi-directional GRU. We also build a Long Short Term Memory Model (LSTM) for text generation, which is able to generate text for each score level. Then we build a recommendation system based on Latent Factor Model, with Stochastic Gradient Descent, and recommend 10 items to selected users. Finally, we use softmax regression to visualize the most important words for a certain score, and design a spam review binary classification based on the helpfulness scores of the reviews.