Online Learning Algorithm Research For Multi-Label Classification

In multi-label classification, any single sample could be related to several labels simultaneously and the class labels are no longer mutually exclusive. At present,multi-label classification has been widely used in numerous domains, such as text categorization, scene classification, gene function prediction and music emotions annotation. Thus a significant number of multi-label classification algorithms have been proposed. In recent years,most of the multi-label classification algorithms are batch learning, which require that the whole training data set is read into memory and is processed for the decision model at once. However, in practice, these batch learning methods, especially for the classification of large-scale data sets, will consume a lot of time and space resources. Aiming at these issues, based on the online learning theory, two kinds of multi-label online classification algorithms are proposed to deal with large-scale multi-label classification problem.1. Via combining binary relevance approach with the existing online binary Passive Aggressive Active learning algorithm, a multi-label online Passive Aggressive Active learning algorithm named MLPAA is proposed. The algorithm uses active learning to query the multi-label sample information. It not only utilizes the method of online learning to update the multi-label classifier model,but also uses the active learning to explore the information of the unlabeled sample, which reduces the human cost and time. In the experiment, according to five multi-label evaluation measures, I compare MLPAA algorithm with three algorithms (MLRPE, MLPEA and MLPEA) on eight multi-label data sets. The experimental results demonstrate that the MLPAA algorithm has better classification performance than other multi-label online algorithms.2. Considering the correlation among labels based on the idea of label order and evolutionary multi-class online Passive Aggressive classification algorithm, I proposed the multi-label online Passive Aggressive classification algorithm by considering the correlation among labels (MLLRPA), which builds a sorted error set by predicting the pairs of labels by maximizing the margin between the relevant and irrelevant labels in multi-label samples. The MLLRPA algorithm updates its classifier model according to the size of error set. In the experiment, I compare MLLRPA algorithm with other three algorithms (MMP, BR-PE and BR-PA) on 10 multi-label data sets, according to four popular and indicative performance measures. The experimental results show that the MLLRPA algorithm has a good performance.