Dual Set Multi-Label Learning

In machine learning,multi-label learning is an important research area where an object can be associated with multiple labels.Compared with single-label learning,the output space of multi-label learning grows exponentially in the number of labels,which leads to great difficulties of the learning task.Therefore,exploiting label correlations is shown to play a key role in improving prediction results.In many real-world multi-label learning tasks,the label space is composed of two label sets with dual relationships.For example,in calligraphic classification tasks,one needs to simultaneously predict calligrapher labels and font labels.Here,each callig-rapher or font is a class label and all calligraphers and fonts are two label sets where an instance is associated with only one positive label in one label set.Although tradi-tional multi-label learning methods can deal with this kind of problem,they ignores the explicit inter-set and intra-set label relationships,which makes them hard to achieve optimal performance.The thesis names this kind of problem as dual set multi-label learning and mainly includes the following research achievements:(1)At the first time dual set multi-label learning is formulated and defined,where the label space is composed of two label sets.The labels within each label set are exclusive and the labels between label sets are in dual relationship.These explicit label relationships make each instance associated with only one positive label from each label set,which are helpful for solving this kind of problem.Besides,three potential approaches are proposed in this paper and theoretical analyses are given to show it is better learning from dual label sets than directly learning from all labels.(2)An efficient algorithm named DSML is proposed specifically for dual set multi-label learning problems.It exploits label exclusive relationships within each label set with base learners and captures dual label relationships between label sets by model-reuse and distribution adjusting mechanisms.These two mechanisms are able to make label sets help each other.In addition,experiments show that DSML algorithm signifi-cantly improves the predication accuracies and the diagnostic test shows the effective-ness of the mechanisms.(3)A calligraphic image classification system is developed based on dual set multi-label learning.Given certain memory and storage of a smart phone,the system is able to classify calligrapher and font labels simultaneously with image collection and labeling,instance feature extraction and classifier training,and image predication steps.