Information Cascades in Yelp Reviews
https://arxiv.org/pdf/1712.00903.pdf
With the information age, more people are able to see the decisions and thoughts of others on events, places, and services. This extra knowledge can easily create biases in peoples own opinions and form information cascades. On platforms like Facebook and Yelp, reviews could potentially have an high impact on how people new to the service/business initially perceive the that service/business. And similar to the situations that we have discussed in class, this might cause people to ignore their personal information on the service/business. Understanding the features behind the cascade and being able to systematically predict trends can help business owners understand the state of their own business.
The attached paper written by Muhammad Khan specifically looks at Yelp reviews, with each city as stand alone systems. In this work, Khan attempts to understand and predict cascade growths using a few initial reviews and determine the features of the nodes and businesses that have high predictive power. To simplify the study, Khan assumes that only the friends of the person writing a review is part of his or her social group and categorizes cascades as big if their length if greater than ninety percentile of the lengths of all the cascades of that city. He then runs a classifying algorithm to track the features (number of reviews and friends, yelp age, gender, and average review of the user) of both the root (or cascade origin) reviewer and the non-root reviewers.
The study was able to predict the longevity of the study much more accurately for the cities with more businesses and reviews, up to 97% for Pittsburgh. It makes sense that smaller systems have more fragile cascades since the individual reviews would be separated by large time gaps. The study was also able to conclude that the properties of the non-root reviewers were more impactful on the cascade growth than the root reviewer, which again is very much a valid conclusion as there are much more non-root reviewers who are influenced by their personal experiences rather than other people’s reviews. And finally, the study was able to determine that the variance in age of the non-root reviewers was the most important feature in most cities. This is perhaps the most surprising of the conclusions, as the credibility of the reviewer is heavily based on how well received his or her past reviews are. However, at the same time it makes sense that if an opinion is reoccurring throughout all age groups then that opinion would be appealing to all age groups in the network, giving it more chances to spread.