Department of Computer Science
Currently my research focuses on learning from structured data. It is often the case that data is most naturally represented structurally, using a graph, for instance. Attempts to propositionalize the data (that is, turn it into a feature vector) make it very difficult to capture the important information contained in the structure. A very popular example of this phenomenon is learning from chemical structures. In this setting, the problem is to predict the behavior of molecules, i.e. their carcinogenicity. Using detailed domain-specific knowledge we can determine the important characteristics of the molecule and give that information to the learner. The problem with this approach is that the bias imposed on the learner by the domain expert may (and in practice seems to) eliminate critical information. Furthermore, there are many domains in which the important characteristics of the data cannot be easily represented by propositional features. Given these challenges, the development of a learner that leverages the important features of structured data with minimal interference from domain experts remains an important open research area. Recent approaches have taken the form of so-called "graph kernels," which are functions that take as input graphs and give as output a real-valued score that allows a learner to differentiate between different kinds of graphs. The construction of such a function is tricky, however, and while such approaches have proved promising, they have not been shown to significantly out-perform existing methods in general.
A second major research area is understanding the effects of margin on learning algorithms. In previous research on noise-tolerant perceptron variants, we discovered that a variant built for finding a large-margin separator performed better in noisy environments than variants built for noise tolerance (i.e. with a L1/L2 slack term). Current learning theory tells us that the upper bound on the error rate of a learner is inversely proportional to the margin, but it falls short of telling us why margin is a "good thing." Looking at the Perceptron algorithm in particular, current proofs of the bounds on various Perceptron variants typically penalize the learner for making margin mistakes. An important contribution to the community would be a proof representing margin updates as improvements.
e-mail address: uid: gwachm01 domain: cs.tufts.ee-dee-you