Learning Models of Users’ Informational Needs

Most current models of users’ informational needs capture only the content, e.g., search terms, and ignore other aspects, such as metadata. However, a recent study has shown that in some cases, performance suffers dramatically when these additional aspects are ignored. But how can more complex user models be built? Motivated by prior work, we explore using machine learning methods to learn the weights of the various aspects, which we call criteria. Of particular interest is whether the criteria combine in a nonlinear way. We evaluate the use of linear and logistic regression on a criteria-based news filtering data set. We also present a way to learn nonlinear models with linear regression, and propose a new loss function based inspired by an common information retrieval metric.