Analysis of Kiva Lending and Team Network Structure

This document describes the techniques used in and results of our project to apply social and information network analysis to the Kiva.org online microlending network. The Kiva network can be modeled as a set of bipartite graphs linking lenders to loans, lending teams to loans, and individual lenders to lending teams. Folding the lender-loan graph to create a graph of lenders linked by common loans results in a network that reflects the connections lenders have based on their own lending preferences. Prior studies have shown that lenders who are part of lending teams are 20% more productive in funding loans than those who are not. This project accessed data from Kiva.org and used personalized page rank vectors (PPRV) from a random walk with restarts algorithm and comparison with nearest neighbors in the folded lender-lender graph to recommend teams similar to a target user. The standard metrics of precision and recall were used to evaluate the methods against users who are already members of lending teams. Average recall rates were 35-51% depending on the technique used and the number of teams recommended to the user with the highest average recall being 51% for a weighted PPRV voting technique to recommend the top 20 teams. The highest average precision was 12% using a nearest neighbor voting technique to recommend the top five teams.