Imbalanced Networked Multi-label Classification with Active Learning

Abstract

With the rapid development of social networks, the networked multi-label classification algorithms have gained wide attention. The existing networked multi-label classification algorithms mostly only consider the homogeneity or heterogeneity of the network without taking the imbalance of the network into account, and this is actually pretty common in real network environments, which deserves more attention. Moreover, the selection strategy of training set is very critical for multi-label classification algorithm, because it will directly affect both the parameter updating inside the classifier and the precision of the classifier. The application of active learning to the selection of training set can effectively improve the precision of the classifier. Similarly, the application of imbalanced data processing strategies to the selection of training sets also makes classifiers more suitable for imbalanced data networks. Thereout, we propose an algorithm BSHD (Block Sampling with selecting the Highest Degree nodes), which is an active learning based imbalanced networked multi-label classification algorithm. In this algorithm, we divide the network according to the edge density and utilize the oversampling and undersampling to dispose each block. Then we select the nodes with the highest degree from each block to form the training set. Experimental results show that our proposed BSHD outperforms other state-of-arts approaches.