Link Prediction in the Yelp Social and Review Networks

In this project, we apply two predictive network algorithms to the Yelp academic challenge dataset. The first algorithm is adapted from Leskovec, Huttenlocher and Kleinberg (2010) to predict the rating that a user will assign to a business [1]. We adapt the notion of signed edges in the network to account for both useruser edges (friendships) as well as user-business edges (reviews). A logistic regression model is trained from features in the social and review network data, and experiments are carried out using ten-fold cross validation. We find that the ability to predict a user’s rating of a business depends highly on the prior ratings given by the user and the prior ratings received by the business. While a user’s rating is impacted by the ratings given by his or her friends, this influence appears to be a second-order effect. In addition, the logistic regression classifier achieves a moderate improvement over a baseline classifier. The second algorithm is adapted from Backstrom and Leskovec (2011) to predict the friends that a user will link to in the future [2]. We extract eleven features from the Yelp dataset and separate friendships into training and testing sets. The training edges remain in the social network; the testing edges are removed. We then optimize a personalized weight vector for each user with gradient descent using a supervised random walk. We find that 56% of the testing edges are ranked in the top 20 recommendations for 300 randomly selected users. Because the candidate set can contain thousands of possible edges, this result is moderately significant. The applications of these predictive models in the Yelp data are twofold: the first algorithm enables a user to predict their future evaluation of a business; equivalently a business can search for potential new clients. The second algorithm enables a friendship recommendation engine, which improves the user experience of the Yelp service, which benefits both users and businesses.