Prediction of Yelp Ratings Based on Reviewer Comments Segmented by Business Type

Yelp[1] is a popular website that allows users to submit numerical ratings (integers 1-5, inclusive, 5 being positive sentiment) and text reviews (reviewer’s comments) to express their experiences interacting with various businesses. This information helps others determine which businesses best fit their needs. The prevalence and utility of rating and review services provide the opportunity and motivation to study the prediction of ratings based on reviews. While these services are prevalent, many internet forums are dedicated to publishing user reviews of products such as laptops[3] and vehicles[4] without quantitative ranking. Often, forums hosting qualitative reviews such as "laptop stopped working but helpdesk was useless!" and "customer service was horrible" would benefit from an algorithm predicting rating from reviews to summarize user sentiment quantitatively. Furthermore, prediction algorithms allow rating and review websites gather additional rating information by predicting from external enthusiast forums where reviews are more descriptive. Our goal was to apply machine learning techniques to Yelp’s published dataset to accurately predict ratings from reviews. An input of a string of words into our Multinomial Naive Bayes algorithm would output a predicted rating between 1 and 5 describing the customer’s sentiment regarding a business. While this has previously been attempted[5][6], most approaches do not focus on training data segmentation. We believed that segmenting training data by business sector (e.g. restaurants) allows more accurate predictions because feature sets become less "diluted" or noisy and words gain meaningful value when given a focused context. For example, an auto mechanic business customer may be "extremely satisfied with the new cold air intake system!" while a restaurant customer may say that "the food took a while and came out cold!" Taken together, if feature