According to Carlos André Reis Pinheiro, social networks in communications are easy to understand and detect, so Oi Telecommunications chose that route first when trying to detect fraud.
Community detection for fraud proved to be somewhat different. It is a progressive search, from looking at the entire network to looking at a group of customers and then within those groups to find unexpected behaviors - outliers.
Narrow the field
Pinheiro said that the first step is cleansing the data to remove phone numbers that are constants across the network. Those numbers might include the call center number and extensions from international calls.
Secondly, Oi Telecommunications needed to understand how to group people. “If you look at every small connection, every small call, you will have a huge community,” said Pinheiro. “We needed to understand how to divide people into relevant communities; we needed to define some boundaries.”
Tighten the linkage
According to Pinheiro, large networks, such telecommunications networks, follow a power law distribution, meaning they contain a small number of communities with a large number of nodes and the majority of the communities have only a few nodes. You can change the size of the communities by changing the value of resolutions.
Bigger communities mean more members, but weaker strength in the links that connect them. Conversely, smaller communities have fewer members but stronger links among the nodes. This metric, called modularity, is the average number of distinct connections.
“This may take trial and error to get the right modularity at first to solve your business problem – is it fraud, churn?” said Pinheiro. “My average number of distinct connections is 10, so I’ll try to end up with an average number of members in the communities like 10. With high-performance analytics, you can test many different types of communities.”
Spotting oddballs
In this research, Pinheiro says that Oi Telecommunications collected three months of data. From the data, they could see those customers who seemed to be committing fraud, flag them and follow them through to the next step. “When you talk about social network, we all think about influence. People are influencing others to follow along in some type of event – like churn or a purchase,” said Pinheiro.
The data from the social network analysis show that in the churn field, the viral effect of communities is very real - 11 percent of churners can play as leaders and they can affect 8 percent of the people they are related to.
“When you talk about purchasing, 14 percent of our purchasers can play as leaders and they can influence 17 percent the entire network,” said Pinheiro.
For analysts, the bad news is that there is no viral effect in fraud. Fraudsters create a community with the express purpose of committing fraud. The good news is that there is no viral effect – fraudsters don’t spread their ideas like those who are thinking of changing providers or who’ve found bargain prices. So, according to Pinheiro, he and his team can just change tactics to search for fraud communities.
Pinheiro says that social network analysis now helped them see the differences in the calling patterns within a community. These differences were measured so that outliers became obvious and investigable.
“Fraud is business,” said Pinheiro. “So there is always this behavior to find and investigate.”
Read Pinheiro’s paper, Community Detection to Identify Fraud Events in Telecommunications Networks. Also read Jodi Blomberg's paper, Twitter and Facebook Analysis: It’s Not Just for Marketing Anymore, about catching criminals using social media. (You can also watch as Anna Brown interviews Blomberg about her paper in this Inside SAS Global Forum video interview.)
