Research On Application Of Distance Metric Learning For Multi-Instance Multi-Label Learning

MIMLL(Multi-Instance Multi-Label Learning)is a significant branch of machine learning.In Multi-Instance Multi-Label learning framework,an example is described with multiple feature vectors and associated with multiple labels.Because of the structure complexity and multi-semantic property of images and texts,image and text classification and nearest neighbors searching problem can be defined as MIML problems.Moreover,Hausdroff distance is widely used to measure distance between bags in feature space in traditional MIML algorithms,however,which sometimes may fail to reflect the true semantic relationship between examples.To overcome this issue,distance metric learning can be used to find a semantically consistent distance metric by utilizing the label information of data.We propose a novel distance metric learning method to enhance the classification algorithms and neighbors searching algorithms under the MIML framework.The main contribution of this paper is as follows:We propose a MIML distance metric learning algorithm based on clustering strategy,during the learning procedure of which,the label correlations are also taken into account.By combining clustering strategy,the computational cost of our proposed metric learning algorithm can be reduced and the relationship between instances and labels can be also explored.We proposed a kNN(k-Nearest Neighbors)based MIML classification algorithm combined with our proposed metric learning method,which calculates the nearest neighbors of multi-instance samples using the learned MIML distance metric to take semantic consistency into account and enhance the performance of classification.We proposed an anchor graph based MIML hashing algorithm combined with our proposed metric learning method,which constructs the adjacent matrix between samples and anchors using the learned MIML distance metric to take semantic consistency into account and enhance the performance of hashing.The eventual experimental results verify the effectiveness of our proposed algorithms.