FLIP: Active Learning for Relational Network Classification

Active learning in relational networks has gained popularity in recent years, especially for scenarios when the costs of obtaining training samples are very high. We investigate the problem of active learning for both singleand multi-labeled relational network classification in the absence of node features during training. The problem becomes harder when the number of labeled nodes available for training a model is limited due to budget constraints. The inability to use a traditional learning setup for classification of relational data, has motivated researchers to propose Collective Classification algorithms that jointly classifies all the test nodes in a network by exploiting the underlying correlation between the labels of a node and its neighbors. In this paper, we propose active learning algorithms based on different query strategies using a collective classification model where each node in a network can belong to either one class (single-labeled network) or multiple classes (multi-labeled network). We have evaluated our method on both single-labeled and multilabeled networks, and our results are promising in both the cases for several real world datasets.