Predicting Yelp Review Star Ratings with Language Feature Analysis

For an assignment, we investigate multiple features regarding Yelp reviews in order to construct a predictor for review star ratings. Our supervised learning model uses linear/ridge regression to observe the correlation between a set of features and review star ratings. Basic readily available features include the business’ star rating, the user’s average star rating, and the total number of votes associated with the review. For advanced features, we discovered that some language processing techniques on review text lead to good features correlating with review star ratings. We combine Latent Dirichlet Allocation (LDA) with other optimizations such as stemming and rounding of edge cases to improve upon the basic feature model. We compare the model’s results with a baseline model using the mean squared error (MSE) as the metric. The baseline resulted in an MSE of 1.67836502285. Our model using the features we described resulted in an MSE of 0.732726208483, which is an improvement over the baseline results by %56.342857572