Analyzing Peer-to-Peer Lending Networks Using PageRank and Hubs & Authorities
http://snap.stanford.edu/class/cs224w-2014/projects2014/cs224w-60-final.pdf
The paper above is a final project paper from Stanford’s Analysis of Networks class. The paper explores the properties of an online peer-to-peer money lending network called Prosper, a service that has millions of users and has facilitated over two billion dollars of funded loans. The idea is simple: if you want to borrow money, you post a loan listing and lenders lend money when they find a listing that they are interested in.
A natural way to think about this peer-to-peer lending network is as a directed graph where we consider nodes as users of the service and edges as indications that one user has gotten a loan from another. We can use approaches we learned in class to then analyze the trustworthiness of users. Intuitively, if a user is successfully getting a lot of loans, the user is more trustworthy. We could generate scores through procedures like Hubs and Authorities or PageRank.
The paper above explains that if we direct edges from borrowers to lenders and then use Hubs and Authorities, we can see that users who borrow a lot have higher hub scores, while frequent lenders have higher authority scores. You could just as easily use PageRank to determine the trustworthiness of a user. The paper above found that histograms of hub scores, authority scores, and PageRank scores all followed a power law distribution. As a result, the top lenders and borrowers were responsible for most of the transactions on the platform.
The paper then also used the PageRank, hub, and authority scores as some of the features to various machine learning models that were trained to predict the mode of a user’s credit level over time. The authors of the paper trained a random forest model that had an F-measure of 0.72 on the task, which suggests that the PageRank, hub, and authority scores have some utility in this task.
It’s brilliant how techniques initially envisioned for web search can be applied in such interesting contexts!