Computer Science Department
School of Computer Science, Carnegie Mellon University


New Event Detection with nearest Neighbor,
Support Vector Machines, and Kernel Regression

Jian Zhang, Yiming Yang, Jaime Carbonell

April 2004

This document also appears as Language Technologies Institute
Technical Report CMU-LTI-04-180

Keywords: Artificial Intelligence: Learning, Pattern Recognition: Models-Statistical; Pattern Recognition: Design Methodology-Classifier design and evaluation; algorithms, novelty detection, new event detection, nearest neighbor, support, vector machines, kernel regression

Support Vector Machines have received extensive attention in machine learning community and have been successfully applied in pattern recognition and regression problems. Recently, it has also been proposed to solve novelty detection problems, whose objective is to detect novel objects from existing instances. New Event Detection (NED), which can be treated as one special application of novelty detection, has been a research topic in Topic Detection and Tracking (TDT) community for several years. However, the winning technology of NED in the TDT community has remained to be the nearest neighbor method with suitable distance metric in the document vector space. In this paper we investigated Support Vector Machines and kernel regression (as a smoothed nearest neighbor method) for the NED task, and compared them to the nearest neighbor method. We conducted a set of experiments on TDT benchmark collections, and provided analysis on the failure of SVM for not being able to capture misses.

36 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by