[:is] Math colloquium
Fyrirlesari: María Óskarsdóttir, Háskólinn í Reykjavík
Titill: Ranking nodes relative to influence with the Personalized PageRank algorithm applied to fraud detection and credit risk measurement
Staðsetning: VRII-258
Tími: Fimmtudagur 12.mars kl. 10:50
Ágrip:
Various phenomena in both the physical and the digital world can be represented with networks, that is, entities that are connected in some way, for example communication, computer, financial and social networks. A central theme in the analysis of networks is finding the most important nodes in a network. The PageRank algorithm was developed to rank webpages in search engines, to find the most important webpages on the internet, but has been applied in numerous others applications. The ranking can be personalized so that nodes which are important relative (or close) to a predefined set of nodes are ranked higher. This approach has been used to identify certain behavior in networks where there is a strong social effect, for example fraud and churn. In this presentation we show how the personalized PageRank algorithm can be extended for two specific types of networks. First, we look at a bipartite network which consists of claims and the involved parties, i.e. policyholders and brokers, with the goal of finding fraudulent insurance claims. Then we consider multiplex networks, in which each node can be connected to another node by more than one type of edge, such as two different networks connecting the same individuals. They arise naturally in lending, as two borrowers can be connected by geographical location, economic activity, and many other relationships. We present a methodology to leverage multiplex networks by a novel multiplex Personalized PageRank algorithm, which we subsequently apply to credit risk assessment. [:en]
Speaker: María Óskarsdóttir, University of Reykjavík
Title: Ranking nodes relative to influence with the Personalized PageRank algorithm applied to fraud detection and credit risk measurement
Room: VRII-258
Time: Thursday 12th March, 10:50hrs
Abstract:
Various phenomena in both the physical and the digital world can be represented with networks, that is, entities that are connected in some way, for example communication, computer, financial and social networks. A central theme in the analysis of networks is finding the most important nodes in a network. The PageRank algorithm was developed to rank webpages in search engines, to find the most important webpages on the internet, but has been applied in numerous others applications. The ranking can be personalized so that nodes which are important relative (or close) to a predefined set of nodes are ranked higher. This approach has been used to identify certain behavior in networks where there is a strong social effect, for example fraud and churn. In this presentation we show how the personalized PageRank algorithm can be extended for two specific types of networks. First, we look at a bipartite network which consists of claims and the involved parties, i.e. policyholders and brokers, with the goal of finding fraudulent insurance claims. Then we consider multiplex networks, in which each node can be connected to another node by more than one type of edge, such as two different networks connecting the same individuals. They arise naturally in lending, as two borrowers can be connected by geographical location, economic activity, and many other relationships. We present a methodology to leverage multiplex networks by a novel multiplex Personalized PageRank algorithm, which we subsequently apply to credit risk assessment. [:]