Identifying Influential Users in Social Network with Review Data

Background: Social networks have been widely utilized by a variety of popular review websites, and different users have a different impact on how information propagates through the network. Identifying influential users in such system could be beneficial for advertising purposes. Aim: To estimate the influence propagation probability for social ties in the network using review data which contains rating and temporal information. Then to identify influential users in the social network. The new method should be able to tolerate some basic attacks. Data: The Yelp dataset was used in this analysis, and we focused on a subset of 5691 users who have at least one of 40077 reviews about 2724 businesses in Pittsburgh. Synthetic dataset for attacks is generated on top of the Yelp dataset where up to 10000 fake social ties are added in the network. Method: The convex network inference model was adopted and applied the algorithm with prior knowledge of the social network. We then used a new method combining the two modifications on top of this method: the self-exploring model which supports user’s information discovery without knowledge propagation; the rating model which incorporated rating data. The credit distribution model for influence maximization was used for comparison. Result: Our new model achieved the best outcome in finding a set of influential users that have a high influence on the network. We also simulated attacks on the social network which changed 0.55% to 55% of the social network. Our results show that the total weight of the entire network can be changed up to 80%, but the influence per review for a set of the 100 most influential users only changes by 18% with this attack. If attackers focus on using users with only high number of reviews, this change could go up to 26.6%. Conclusion: The network inference model is adopted and modified to estimate the influence propagation probability with review data, which is shown to perform well in identifying influential users in the network while being able to tolerate attacks on the social network without a severe impact on the result.