The Research On Multi-label Classification Algorithms Via Nonnegative Matrix Factorization

Multi-label classification algorithm is a research hotspot in data mining technology.Multilabel classification algorithm can be used to solve the problem of classifying the sample data with multi labels.The research results are widely used in various fields such as text information classification,semantic annotation of image video,functional genome and music emotion classification.Compared with the traditional single-label classification algorithm,which belongs to only one category label,the multi-label classification algorithm can satisfy the actural needs of practical application,because one sample in the practical application belongs to the multi-label data types which is more likely to have multiple attributes at the same time.In this paper,MLNMF and i-MLNMF are proposed to solve the multi-label classification problem.I-MLNMF is the extension algorithm of MLNMF.I-MLNMF algorithm will be more effective in dealing with large data,because the probabilistic formula of i-MLNMF is changed which is based of the original MLNMF algorithm.Several multi-label classification algorithms that have been proposed.Although these methods will be able to classify the multi-label samples,the problem of how to make the samples classified to corresponding labels more efficiently and how to mine the corresponding relationship between the sample characteristics and the labels is still not completely resolved.Compared with the proposed algorithm,our method can guarantee the results of multi-label prediction and reduce the running time of the algorithm.Both MLNMF and i-MLNMF are utilize the NMF method to generate the corresponding relationship between the sample characteristics and the labels,the label probability prediction model(LPPM).The classifiers of labels will be generated by a decision stump method.Finally,the model and the classifier are used to predict the labels corresponding to the unlabeled samples.The experimental results show that the MLNMF and i-MLNMF algorithms are better than or close to the classical algorithm especially on large data sets and achieve higher performance in terms of the efficiency of the algorithm.Therefore,our methods have very high practical value in the actual needs of multi-label classification.