一种基于构建2竞争聚类及KNNFL 的事件探测与追踪系统

一种构建-竞争聚类法被用于事件探测,该方法是受神经网络研究中构建-竞争学习的思想启发的.另外,提出了一种用于事件追踪的基于K近邻特征线(KNNFL)的分类方法,这种基于最近邻特征线(NFL)的方法本质上可以看作是对K近邻(KNN)法的推广,将改进后的KNN融入到NFL中形成KNNFL是为了更适合新闻事件的分析.研究结果表明,本文所提出的方法与传统的增量k均值法、Single-Pass法、Rocchio法以及KNN法相比较,可以获得更好的效果.通过分析可以看到,KNNFL即使在正例样本非常稀少的情况下仍然具有鲁棒性的表现.

A System for Event Detection and Tracking Based on Constructive-Competition Clustering and KNNFL

The objective of event detection and tracking is to automatically spot previously unreported new events from news-feed and assign documents to previously spotted events.A Constructive-Competition Clustering(C3) method was used for topic relevant event detection in this paper,which is motivated by constructive and competitive learning from neural network research.In addition,a classification method based on K Nearest Neighbor Feature Line(KNNFL) was proposed for tracking events,this method based on Nearest Feature Line(NFL) is essentially an extension of the K Nearest Neighbor(KNN) method,NFL combining with improved KNN produces KNNFL in order to make it more suitable to news event analyzing.The study indicates that the proposed methods in this paper achieve superior performance than the traditional incremental k-means,Single-Pass clustering,Rocchio and KNN.The computational analysis has showed that,KNNFL behaves robustly even if the number of positive training examples is extremely small.