Research On Multi-instance Learning Based On Support Vector Machine

The growth of science and technology has stimulated the emergence of the era of big data,which makes data has step-by-step becoming a significant resource to promote the development of all industries and business functions today.How to mine potential and valuable information from exponentially growing data has become an issue of widespread concern in society.Multiple instance learning means that each data sample(i.e.bag)in the training set contains multiple instances,and only the complete label information of the bag,the label information of the instance is incomplete.However,in addition to the lack of effective measures for the correlation between instances,there is also an imbalance in multiple instance learning.That is,multiple instance data set are not evenly distributed.Therefore,this paper conducts an in-depth study of multiple instance learning based on the above two problems,the following is the main research work:Based on instance features and instance correlation,this paper proposes a multiple instance learning algorithm,which name gray-based multiple instance learning with multiple bag-representative(MIMBR).For each bag in the data set,it integrates the instances in the bag through gray correlation analysis,and makes a preliminary judgment on the importance of the instances in the bag.In order to simplify the decision function and reduce the computational complexity of the decision function,this paper trains a two-step iterative support vector machine optimization framework for updating classifiers,which based on bag-level information.In order to expand the wide applicability of the algorithm,the multiple instance learning algorithm is extended from binary classification to multiple classification.In order to verify the effectiveness of the algorithm,this paper evaluates the performance of the MIMBR algorithm with 9 state-of-the-art multiple instance methods on 10 data sets,the results confirm that MIMBR algorithm has better classification performance.Considering how to better improve the problem of class imbalance,this paper proposes a method to improve the positive class representation in multiple instance learning.It can effectively integrate the instance information in the positive bag to build a classification model or classifier.In multiple instance learning,the distribution of instances in different data sets are different,and the kernel density estimate can reflect the overall distribution of negative instances from the side.This algorithm according to the distribution of instances,for each instances in the negative bag,each instance in the negative bag is given a weight in descending order according to the number of positive instances in the neighborhood.This can ensure that the instances in the positive bag can be classified as correctly as possible.After using the processed positive bag and the negative bag selected by weight as a new data set,the experimental results show that the class imbalanced problem in multiple instance learning can be effectively solved.This paper proposes two multiple instance learning algorithms based on instance features and correlations,and their effectiveness has been fully proven through a series of theoretical analysis and experimental verification.Finally,a brief summary of the main work and innovations made,and make an outlook and plan for the next research.