Researcher teaches computers to detect spam more accurately
- 11 August, 2011 08:08
Without spam detection many of us would spend hours managing the daily load of e-mails. Nina Balcan develops machine learning methods that could be used to develop personalized automatic programs for deciding whether an e-mail is spam or not. For her efforts, the computer scientist from the Georgia Institute of Technology has just been awarded with a Microsoft Research Faculty Fellowship.
Balcan is one of eight recipients of the fellowships the company has awarded this year. As a fellow she receives US$200,000 to be used at her discretion. According to Microsoft the total funding for the program is $1.4 million.
The results of Balcan's research can be broadly used to solve many data mining problems, she said; spam detection is just one of many possible applications. As to a computer's decision if an e-mail is a true invitation to Facebook by one of your friends or just a message sent by someone phishing for personal data, there are several methods to teach the machine how to decide.
With so-called supervised learning, the user teaches the computer by feeding it with information on which e-mails are spam and which are not. "Therefore a person has to label a huge number of mails as spam or not spam, which is not the most efficient," Balcan said.
A better method would be to use active learning, she said. With this method the computer makes use of huge collections of unlabeled e-mails to adaptively generate only a few questions for the user. This method reduces the effort the human user has to perform, as he or she only needs to label a handful of e-mails. Typically, the questions would be about the unlabeled examples the computer is most uncertain about given its current knowledge. The computer can identify patterns in e-mails the user has already labeled. Then, it selects mails from the unlabeled pool that do not match these learned patterns or are borderline. By using these uncertain cases the computer learns new patterns, Balcan said.
Active learning has the potential to always deliver better results than supervised learning, Balcan said. However, active learning methods are often highly sensitive to noise, making this potential difficult to realize. A single non-spam e-mail labeled as spam by mistake could send the system in a bad direction from which it would never recover. The aim of her work is to change this. "I am looking for robust methods," she said.
More generally, Balcan's aim is to develop an understanding of when different kinds of learning protocols help, why, and by how much. Her approach is interdisciplinary. "My research connects machine learning, game theory, economics and optimization, " she said.
Balcan collaborates with statisticians, theoretical computer scientists, game theorists and computational biologists. The Microsoft fellowship could boost this interdisciplinary work. Balcan says she will use the money mainly for more interaction with other researchers. She says she is in touch with many researchers from all over the world. The money makes it easier for her to invite them from time to time or travel to them for collaboration. Moreover, she is going to hire postdoc researchers and graduate students for her research group.
Balcan previously worked with Microsoft from October 2008 to July 2009, when she was a postdoctoral researcher at Microsoft Research New England in Cambridge, Massachusetts.
She joined the Georgia Institute of Technology as an assistant professor in July 2009. Balcan finished her Master of Science in computer science at the University of Bucharest in Romania and received her Ph.D. in computer science in 2008 from Carnegie Mellon University in Pittsburgh.