Active semi-supervised learning using particle competition and cooperation in networks

Both Active Learning and Semi-Supervised Learning are important techniques when labeled data are scarce and unlabeled data are abundant. In this paper, these two machine learning techniques are combined into a new nature-inspired method, which employs particles walking in networks generated from the data. It uses combined competitive and cooperative behavior in order to possess nodes of the network, and thus labeling the corresponding data items. Particles represent labeled nodes, and new particles can be added on the fly to the network as the result of queries (new labels). This built-in mechanism saves a lot of execution time comparing to active learning frameworks, since only nodes affected by the new particles are updated, i.e., the algorithm does not have to be executed again for each new query (or new set of queries). The algorithm naturally adapts itself to new scenarios, i.e., more particles and more labeled nodes. Experimental results on some real-world data sets are presented and the proposed active semi-supervised learning method shows better classification accuracy than its only semi-supervised learning counterpart when the same amount of labeled data is used. Some criteria for selecting the rule to be used to choose data items to be queried are also identified.