|
Abstract:
|
In this talk, we present a general framework that combines feature-based (X) data and graph-based (G) data for prediction of the response (Y). We examine a model fitting approach to fit eta=Xb+f(G) with link eta=g(u), coefficient b, and f is a function over the vertices of the graph. The procedure is semi-supervised in nature (trained on the labeled and unlabeled sets), requiring iterative algorithms for fitting this model. We discuss convergence properties of this procedure and provide a generalization to local scoring algorithms in semi-supervised learning. Applications in text data, proteomics and genomics are discussed. Specifically, in text problems the observations are papers, Y is the paper's category, X is the journal that the paper was published, and G is the co-citation network (vertices represent papers, and edges are citation counts between two vertices).
|
