Research On Multi-Label Classification By Exploiting Label Correlation

In multi-label learning problem, each example is associated with a set of labels simultaneously. The main difference between multi-label and traditional classification is that the labels in traditional classification are assumed mutually exclusive while the labels in multi-label classification are relevant. Obviously, in practical applications, multi-label classification is more general.As the labels in multi-label learning are often correlated, it can improve classification performance to explore the label correlations.Most of the existing approaches exploit labels coexistence based on a tree or graph structure among labels.However, in most multi-label applications,labels are only provided without such structures. In addition,these methods may not adequately depict certain implicit exclusive relationships:sets of labels that do not coexist in the same examples.Learning from multi-label data has received a lot of attention from the machine learning and data mining communities in recent years. This is partly due to the multitude of practical applications it arises in, and partly due to the interesting research challenges it presents, such as exploiting label dependencies, learning from rare labels and scaling up to large number of labels.This work presents a novel multi-label method based on label correlations.The main idea is that those mutually exclusive relationships are represented with distance information among labels by using Euclidean distance.And We proposed several methods with good performance based on label exclusion. The main work this paper has finished is as follow:l.an improved RAKEL algorithm:The RAKEL algorithm constructs a small random subset of labels and learn a single-label classifier,it cannot fully take advantage of label correlations.Aimed at this disadvantage,in this paper we choose a mutually exclusive labels to constructs a small subset.Experimental results show that our algorithm can achieve better performance.2.an improved CC algorithm:The CC algorithm constructs a label chain of binary classifiers by using random order, it also cannot fully take advantage of label correlations.Aimed at this disadvantage,in this paper we choose a mutually exclusive labels to constructs a small subset.Experimental results show that our algorithm can achieve better performance.