Introduction
This article seeks to summarize the article Amazon Food Review Classification using Deep Learning and Recommender System by Zhou and Xu. The main aim of this article is to analyze, the problem trying to be solved by the researchers, the techniques or the functions from deep neural networks used and the implementation of the recommendation system used by Zhou and Xu (2016).
What are the problems they are trying to solve?
The main problem being solved is the review usefulness in the classification problem. According to Zhou and Xu (2016) majority of the ecommerce websites such as the Amazon enable users to write a review on the services and products they purchased during their shopping. The reviews are often critical to other users when trying to make the decision on whether to buy or not to buy the product. This shows that, websites should seek to understand the meaning of reviews and classify them correctly. This way, the classification results can be used in creating an effective recommender and summarization system. Zhou and Xu (2016) classified the usefulness of each review by making use of the deep learning models.
What functions or techniques from deep neural networks are they using?
For this research, Zhou and Xu2016 used the feed-forward neural network and the LSTM model as the main techniques and functions from the deep neural networks. The standard feed-forward neural network was first used where the input was the word vector which represented every review x. The input was forward propagated through the deep network where the output yi was a 2-dimensional vector which demonstrated whether the review was useful or not. Experimentation was done with different number of hidden layers hidden units and the non-linear relationship between the number and the levels of the embedding size. This consequently resulted to a five-layer neural network.
So as to explore more an advanced model, the LSTM model was used which was introduced by Hochreiter & Schmidhuber. In the Model, Zhou and Xu (2016) trained the word vector trained their word vectors on an all words in all reviews in a global word-word occurrence matrix as opposed to using the pre-trained word vector. The X representation was then fed into the LSTM model. In this case, the model was used to essentially map the word sequence to a workflow and a class. Class proportion was used in calculating the cross entropy loss. After input was fed through the LSTM model, the yi output is a 2-dimensional vector which is a representation if a review is useful or not useful.
Both the LSTM and the Feed forward neural network were used to beat the baseline model as the loss function was adjusted which helped the models to improve the accuracy in prediction. Nevertheless, the Feed forward neural network outperformed the LSTM model as word vectors were trained in the model and the LSTM processed sequenced of words and therefore it was able to store more information. The classifier worked in a manner in which, there was a high probability that answers with more positive words and longer answers were classified as being useful
How is the recommendation process implemented?
Zhou and Xu (2016) implemented the standard matrix factorization and collaborative filtering system and opposed to the deep neural network based recommender system as their experienced time constraint. Review classification was used so as to establish consistency in the data set, In addition the test/training set was divided as 70/30 and the dataset used. The baseline model for the recommender system was therefore recommending the products which were overall popular. The measures for analyzing the performance of the matrix factorization and collaborative filtering was the RMSE test set. The different recommender systems were analyzed based on their performances where the RMSE attained for the popular baseline model was 1.7372, Matrix Factorization was 1.1198 and Collaborative Filtering 1.4538. This shows that both the matrix factorization and the collaborative filtering outperformed the baseline model and the matrix factorization was the model which performed best. The matrix factorization is one of the most widely used and successful model as it characterizes both the user and item vectors of factors which are secondary from item rating patterns as opposed to similarity measures so as to be able to recommend items to the user.
Conclusion
The main problem being solved is the review usefulness in the classification problem while the feed-forward neural network and the LSTM model as the main techniques and functions from the deep neural networks and the standard matrix factorization and collaborative filtering system and opposed to the deep neural network based recommender system as their experienced time constraint.
Reference
Zhou Z and Xu L (2016) .Amazon Food Review Classification using Deep Learning and Recommender System. CS224d: Deep Learning for Natural Language Processing. StanFord.edu
freelancerresearch 