Multi-Instance Multi-Label Learning Based On Neighborhood Consensus

For the problem of missing labels,label-specific features,label correlations and instance correlations in multi-instance multi-label learning.Firstly,this paper explores the missing labels problem from the perspective of multi-label learning,and then explores the instance correlations from the perspective of multi-example learning,and last the paper proposes an algorithm named NC-MIML(multi-instance multi-label base on neighborhood consensus)for dealing with the problem of missing label and label-specific features by combining the first two problems.The main work and innovations of this paper are as follows:(1)To solve the problem of label correlations,this paper analyzes and proposes an improved multi-label algorithm IMLDGM(Improved Multi-Label algorithm based on Data Gravity Model)from the perspective of multi-label learning.The IMLDGM algorithm takes into account the relationship between the mass of data particles and the size of gravity and the positive and negative correlation of labels based on the gravity model for multi-label learning algorithm.The results of the simulation experiment show that the label correlations affects the effect of the final classification of the classifier.(2)To solve the problem of instance correlations,this paper analyzes and proposes a multi-instance algorithm NC-mi Graph(Neighborhood Consensusmi Graph)from the perspective of multi-instance learning.NC-mi Graph primarily add the neighborhood consensus(i.e.the near-neighbor sets)of each in-package instance to the original feature space of the instance,and then exploring the instance correlations by building an affinity matrix.The results of the simulation experiment show that there is also some influence on the final classification effect of the classifier if the instance correlations and the domain consensus of the examples is taken into account and the neighborhood consensus of the instances are considered.(3)Based on the first two studies,a multi-instance multi-label algorithm NC-MIML based on neighborhood consensus is proposed.NC-MIML algorithm is mainly the first to transform multi-instance multi-label data into single-instance multi-label data through the idea of NC-mi Graph algorithm.Then an incomplete label matrix label is supplemented by using positive and negative label correlations,and so as to find the label-specific features.The results of the simulation experiment show that the final classification effect of NC-MIML is better than that of the comparison algorithm and has some competitiveness compared with other state-of-the-art multi-instance multi-label algorithms.