Research On Similarity-based Two-view Multi-instance Learning Algorithm

In multi-instance learning,the training set consists of a set of bags,and each bag contains several instances.Only the classification label of the bag is known,and the label of the instance in the bag is unknown.For example,in image classification,an image can be cut into different areas.An image can be regarded as a multi-instance bag and an area can be regarded as an instance.If a bag is marked as a positive bag,it contains at least one positive instance.If a bag is marked as negative,all instances in the bag are negative.Multi-instance learning has important applications in image classification and text classification.In addition,some data,such as image and text,can be described from different views.Learning this data from different views is called multi-view learning.In order to further improve the accuracy of image classification and text classification,this paper applies multi-view learning to multi-instance learning,and proposes a similarity-based two-view multi-instance learning algorithm.The main research work of this paper includes:(1)A similarity-based two-view multi-instance learning method(STMIL)is proposed.First,since the label of the instance in positive bags is unknown,we propose a similarity model and a similarity calculation formula.The k-means clustering algorithm is used to reconstruct the bag.Based on the reconstructed bag,the similarity of each instance with respect to the positive and negative classes is calculated.Secondly,we represent the data,such as images and text,in two views,and incorporate this two-view data into the similarity model.Then,a two-view multi-instance learning support vector machine algorithm based on similarity is designed.Finally,the heuristic framework is used to update the similarity of the instances to the positive and negative classes until the algorithm converges.(2)To verify the feasibility of the two-view multi-instance learning method based on similarity,we compare our method with the existing multi-instance learning methods(GMI-SVM,mi-SVM,DD-SVM,Well SVM and PSVM-2V)in the experiments.In the experimental,we first use three clustering methods(k-means,EM clustering and DBSCAN)and two image segmentation methods(Grab Cut and MILCut)to pre-process the image data and text data,and compare our method with the above five multi-instance learning methods.Experimental results show that our method has higher classification accuracy.In addition,we also add different proportions of noise to the input data and test the anti-noise performance of different multi-instance learning methods.Experimental results show that our method is more robust than the existing multi-instance learning methods.