In this seminar we will host local satellite presentation(s) of women involved in data science at Oregon State University which will serve as satellite event for the Women in Data Science 2017 event http://www.widsconference.org/. (More information and live link for the all-day event is at http://www.math.oregonstate.edu/node/13666)
This event can be viewed at youtube link.
Abstract of the talk: We explore and compare some recently proposed penalized approaches in the context of singular-value type decompositions, principal components analysis, and clustering problems. For these problems, we find that simpler approaches produce faster, easier to interpret, and generally more optimal results than the LASSO-type penalized methods that have been proposed and discussed by several authors. We illustrate some issues arising with the LASSO penalty in these cases, including the difficulty of tuning parameter selection and the artificiality and sub-optimality of the criterion that the LASSO imposes. We propose simple methods with more directly interpretable tuning parameters, and demonstrate the superiority of these approaches on simulated and real data. Furthermore, we investigate the interpretability and limitations of the results of any such data-exploration technique, and discuss the relevance of the simple simulated models that are frequently employed in discussions of such sparse techniques to the complicated datasets that are commonly encountered in real examples.