LAGO, SVM and Rare Target Detection

Speaker: Mu Zhu (Statistics and Actuarial Sciences, UW)

I shall describe a few projects in the area of rare target detection. First, I will go over the main ideas of LAGO, a new algorithm we developed that is much more efficient than SVM and KNN for rare target problems. Second, I will discuss a new experiment in which we try to select training observations sequentially in order to build a good SVM. Our main area of application so far has been drug discovery, but the methods developed and lessons being learned are generally applicable to other rare target problems. Other than myself, people involved in these projects include: Wanhua Su, Michelle Zhou, Hugh Chipman and Stan Young.