The science behind social networking

Photo:

Dr Xiongcai (Peter) Cai believes the progress in people to people recommendation will enable people to find other people easily in online social networks.

Further Information:

Dr. Xiongcai cai
Email: xcai@cse.unsw.edu.au
Website: www.cse.unsw.edu.au/~xcai

Despite the growing dominance of social networking sites, the methods used to search for “links” between people in databases are inefficient and based on simplistic formulas.

Dr Xiongcai (Peter) Cai, Research Fellow in the School of Computer Science and Engineering, has been working with a team of researchers to develop more appropriate ways to search for links between people. He says that traditionally, these searches are done through the more familiar “item- to-people” recommenders. For example, Amazon would keep a track of the products its customers have ordered in the past, and therefore recommend other books to them based on these choices. But in this process, the new books don’t get a choice about who they are recommended to. With systems involving people, the receiver has to “choose” as well as the initiator.

“Previously we didn’t really have any suitable existing techniques to apply to people-to-people recommendations,” Peter says. “The recommendations are based on one side of the equation: if someone likes someone. Most on-line social systems have different kinds of approaches and different kinds of directions, but we decided to focus on that links prediction – to really predict whether there is a link from one to the other, not just ‘is Alice interested in Bob?’, but ‘is Bob also interested in Alice?’.”

Peter’s group has developed prototypes that search for two-way or reciprocal matches on a database of about 2 million names and 20 million “links”. “We have developed several methods and have some trials using those methods, so currently we are trying to prepare something to make it into the real world. Surprising findings derived from our results show that user behaviours are often different from users’ explicit preferences in terms of relationship development.”

He gives an example of how current systems, such as LinkedIn or some internet dating sites would recommend people. “Let’s say ‘Alice’ likes ‘Bob’ and ‘Charlie’.

If ‘Denise’ links with Bob, then the system currently recommends she also link with Charlie, based on what was determined by Alice. But our techniques look at both taste and attractiveness of Alice, Bob, Charlie and Denise. If Alice and Denise are interested in Bob and Charlie, that’s not enough. Denise is really interested in Charlie, but Charlie may not be interested in Denise, so it’s not really a match.”

Peter says the research will hopefully enable people and organisations to find other people or organisations easily. “This kind of research is getting popular at the moment and getting more significant. We’ve taken this to industry partners and they’re quite interested in this as well, and they want to get involved.”