Title: Social networks analytics using GOTCHA: Theory and applications
Authors: Bart Baesens - KU Leuven (Belgium) [presenting]
Abstract: A common assumption in analytical applications is that customer behavior is independent and identically distributed, often referred to as the IDD assumption. However, in many real-life settings this assumption is simply not valid. Social network effects between customers, both implicit and explicit, create collective correlational behavior which needs to be appropriately analyzed and modeled. We will start by outlining the architecture of a social network learning environment, consisting of a local model (e.g. a logistic regression model), a relational learner (e.g. a relational neighbor classifier), and a collective inferencing procedure (e.g. Google PageRank). We will then introduce our recently developed GOTCHA method for social network analytics and will illustrate the application thereof in various real-life settings such as churn prediction, credit scoring and fraud detection. It will be empirically shown how GOTCHA allows us to efficiently model social network effects, hereby generating both additional lift and profit compared to, e.g., a flat logistic regression model.