Classification with active learning and meta-paths in Heterogeneous Information Networks

Abstract

Heterogeneous information networks contain multi-type entities and links compared to homogeneous information networks and thus give us richer information about the relationships among the entities. Meta-path is a powerful way to represent complex connections among objects in heterogeneous information networks. Classifying objects in heterogeneous information networks can give us insight about the hidden structure of the network and help us on recommendation, link prediction, community detection, etc. Multi-type entities drive us to utilize class labels of different type entities simultaneously and do multi-task classification. In this talk, I will introduce how to make use of meta-path to improve multi-task classification results in heterogeneous information networks and take advantages of active learning to save human labeling effort for getting training data. Some experimental results will be shown in terms of classification accuracy and active learning effectiveness.