| Abstract: |
Microfinance has known a large increase in popularity, yet the scoring of such
credit still remains a difficult challenge. In general, retail credit scoring
uses socio-demographic and credit data. We complement such data with social
network data in an innovative manner i.e. with fine-grained interest and
social network data from Facebook. Using a unique dataset of 4,985
microfinance loans from the Philippines, we show how the different data types
can predict creditworthiness. A distinction is made between the relationships
that the available data imply: (1) look-a-likes are persons who resemble one
another in some manner, be it liking the same pages, having the same
education, etc. (2) friends have a clearly articulated friendship relationship
on Facebook, and finally (3) the \Best Friends Forever" (BFFs) are friends
that interact with one another. Our analyses show two interesting conclusions
for this emerging application. Firstly, applying relational learners on BFF
data yields better results than considering only the friends data. Secondly,
the interest-based data that defines look-a-likes, is more predictive than the
friendship or BFF data. Moreover, the model built on interest data is not
significantly worse than the model that uses all available data, including the
friendship data. Hence begging the question: who cares about your Facebook
friends when your interest data is available? |