Robert Beverly.
January 2005
Machine learning techniques are increasingly popular tools for understanding complex biological data. Prior research has demonstrated the power of simple statistical clustering algorithms for disease class discovery and prediction. In this work we examine the efficacy of spectral and divisive clustering on gene expression microarray data. In particular we consider simultaneous expression clustering for diagnostically challenging problems such as tumor subclass classification and prediction. We compare spectral and divisive clustering methods against existing cancer classification datasets. Divisive clustering is notably non-parametric, enumerating an estimate of true class count. Using these two clustering methods, we demonstrate a 50-60% prediction error reduction over earlier results.
[ Return to publications ]