Transfer Learning for Network Classification
In this paper, we will study the problem of node classification in networks with the use of transfer learning. In many real applications, it may often be difficult to find a sufficient number of labels on the nodes for the classification process. When typical social networks often have a significant amount of content in the form of text, it is much harder to characterize the nodes in terms of pre-defined properties of interest (or class labels). This is because the nodes in the social network are autonomously created by different entities, who may not label them in a way, which is specific or friendly to particular kinds of applications. The lack of availability of labels on the nodes is analogous to the training data paucity which is often encountered in the classification problem. However, there is often copious availability of text collections, which describe a wide variety of subjects, including the desired properties of interest. Since social network nodes are also often associated with some amount of text content, such collections can be used as a transfer bridge in the learning process. In this paper, we will examine the use of transfer learning from these domains in order to perform collective classification. We will present experimental results, which show the effectiveness of our approach.