Classification Algorithm On Multi-label Learning

Multi-label learning has become one of the hot research topics in machine learning field.Many real-world learning problems fall into the category of multi-label learning. Conventionalmachine learning and data mining methods assume that each instance is associated with onlyone class label within a number of candidate classes. In many real-world problems, however,one object usually belongs to multiple concepts simultaneously. As for this paper, one samplewith several labels, called multi-label problem has attracted lots of our attentions. This thesishas investigated various methods of multi-label classification, transductive multi-labellearning and instance space transformation, whose applications on different kinds of practicaldatasets.We can group the existing methods for multi-label classification into two maincategories. The first is problem transformation methods that fit data to algorithm. Problemtransformation methods transform a multi-label classification problem either into one or moresingle-label classification or regression problems, and then handle the transformed problemsby the standard single-label learning algorithms. The second is algorithm adaptation methodsthat fit algorithm to data. Algorithm adaptation methods extend specific learning algorithms tohandle multi-label data directly.For the problems of multi-label classification, the main contributions of this dissertationare summarized as follows:(1) A new index function for clustering validity is proposed from the angle of samplegeometry, which tries to solve optimization problems, and achieved good performance.Multi-label learning, combining radial basis neural network and K-means clusteringalgorithm, has achieved good effects，but because the number of clusters cannot be welldetermined in advance, an accurate value of the clustering cannot be obtained. This problemwill lead to lower quality clustering, clustering instability, etc. and then affect the stability andthe classification performance of multi-label RBF neural network algorithm. This paper, fromthe angle of sample geometry, employing an index function for clustering validity, tries to findthe optimal number of clusters for each class, then to solve optimization problems. Theoreticalresearch and experimental results have shown that the improved ML-IRBF algorithm caneffectively boost the better performance in terms of the stability and capability ofclassification.(2) A new transductive local correlation method of multi-label classification is proposedin this dissertation, which is to effectively assign multiple labels to each instance using bothlabeled and unlabeled data. In the approach, fuzzy memberships of each instance are calculated based on instancedistribution and local correlation, and a set of multiple labels is assigned to each instancebased on the finally obtained fuzzy memberships. We modify fuzzy c-means by introducingtransductive local correlation information in the membership updating processing, as we thinkthat the labels of an instance is not only determined by the distribution of the data but alsoinfluenced by its local neighbors. The memberships obtained for each instance is used topredict a label set for that instance according to a label prediction rule.(3) A novel multi-label classification approach with instance space transformation isproposed, which use a separate feature transformation algorithm to reconstructed unseeninstances. In this way, the instance and labels can be effectively matching.Multi-label classification deals with ambiguous examples each may belong to severalconcept classes simultaneously. In this learning framework, the inherent ambiguity of eachexample is explicitly expressed in the output space by being associated with multiple classlabels. While on the other hand, owing to the ambiguity of example itself, it is not so easy todetermine label sets for unseen instances. Based on this recognition, we hypothesize that if theunseen instance can be projected into the space of training instances with known labeled sets,the label sets of unseen instances will become explicit. We justify this hypothesis byproposing a novel multi-label learning approach named ML-IST. Experiments on threereal-world multi-label data sets show that ML-IST achieves highly competitive performancewith other well-established multi-label learning algorithms.