Learning for graphs with annotated edges

Abstract

Automatic classification with graphs containing annotated edges is an interesting problem and has many potential applications. We present a risk minimization formulation that exploits the annotated edges for classification tasks. One major advantage of our approach compared to other methods is that the weight of each edge in the graph structures in our model, including both positive and negative weights, can be learned automatically from training data based on edge features. The empirical results show that our approach can lead to significantly improved classification performance compared to several baseline approaches.