Join us for a research seminar on how social networks can be used in predictive modelling, and how these can be applied in identifying fraudulent insurance claims.

Calendar
Wednesday 4 March 2020
Clock
14:30–16:00
Navigate
LT1A
Microphone
Dr María Óskarsdóttir, Assistant Professor, Reykjavík University

Overview

Unfortunately this event is now cancelled. Apologies for any inconvenience.

This presentation will cover how social networks can be used effectively and efficiently for predictive modelling. Three practical applications will be targeted: churn prediction in the telecommunication industry, credit scoring, and fraud detection in insurance. In the first two cases, networks are constructed using call detail records (CDR) which provide an accurate representation of people's behaviour and have therefore become a great source of data for research in disciplines such as physics, sociology, epidemiology, transportation, and networking.

For churn prediction, the performance of the network learning techniques is compared to that of regular binary classifiers (such as logistic regression and random forests) with features extracted from the network, which is a more traditional approach. The results show that churn influence does not spread far in the social network and churn status within a customer's ego-net is highly predictive of churn.

In the second application, credit scoring, I discuss the added benefit of using mobile phone data, or CDR, in credit risk modelling. According to the results, features representing calling behaviour are most predictive in terms of both statistical and economic model performance. These results have important regulatory, privacy, and ethical implications.

Finally, for detection of fraudulent insurance claims, the goal is to find groups of collaborating fraudsters by linking together claims and the involved parties in a massive social network. As such, we are able to look beyond the classical properties of the claim, the policyholder, and the policy, and study the social structures of fraudsters in insurance fraud detection tools and models.