Online 61. Directed learning [electronic resource] [2012]
 Kao, YiHao.
 2012.
 Description
 Book — 1 online resource.
 Summary

In machine learning, it is common to treat estimation of model parameters separately from subsequent use of the model to guide decisions. In particular, the learning process typically aims to maximize ``goodness of fit'' without consideration of decision objectives. In this dissertation, we propose a new approach  directed learning  which factors decision objectives into the model fitting procedure in order to improve decision quality. We develop and analyze directed learning algorithms for three classes of problems. In the first case, we consider a problem where linear regression analysis is used to guide decision making. We propose directed regression, an efficient algorithm that takes into account the decision objective when computing regression coefficients. We demonstrate through a computational study that directed regression can generate significant performance gains, and establish a theoretical result that motivates it. This setting is then extended to a multistage decision problem as our second case, and we show that a variation of directed regression, directed timeseries regression, improves performance in this context as well. Lastly, we consider a problem that involves estimating a covariance matrix and making a decision based on that estimate. Such problems arise in portfolio management among other areas, and a common approach is to employ principal component analysis (PCA) to estimate a parsimonious factor model. We propose directed PCA, an efficient algorithm that accounts for the decision objective in the selection of components, and demonstrate through experiments that it leads to significant improvement. We also establish through a theoretical result that the possible degree of improvement can be unbounded.
 Also online at

Special Collections
Special Collections  Status 

University Archives  Request via Aeon (opens in new tab) 
3781 2012 K  Inlibrary use 