A Study On Algorithm For Multi-instance Learning Based On Support Vector Data Description

In the1990s, Dietterich etc. firstly proposed multi-instance learning concept which was originated from their research of drug activity prediction. So multi-instance learning has become a new theory in machine learning. As multi-instance learning has unique characters and well application prospect, multi-instance learning is called the fourth machine learning framework with supervised learning, unsupervised learning and reinforcement learning, and it has caused the extensive concern to many researchers. Many researchers studied multi-instance learning’s unique characters, and proposed some methods to solve the problem of multi-instance learning. However, because of the ambiguity of positive bags that we only know there is at least one positive instance in a positive bag but we don’t know which one it is. Other researchers also proposed some methods to solve multi-instance learning problem. In this paper, we analyzed these methods’defects and proposed two new multi-instance learning methods based on SVDD:MI-NSVDDI method and MI-NSVDD B method.This paper studied the following:Firstly, multi-instance learning was reviewed in this paper. Then introduced the support vector data description algorithm and its basic principles and solutions are analyzed and discussed. Then we introduced concept of Support Vector Machine and Support Vector Data Description, and derivation the source of Negative Support Vector Data Description carefully. And we introduced two methods to solve multi-instance problem:mi-NSVDD method and MI-NSVDD, and then we analyzed these methods, and present these methods’ defects.Secondly, this paper proposed a disambiguation method to convert the sample data set to a single sample data set. We calculated all instances’prediction accuracy, and then we selected an instance which has largest prediction accuracy in each bag. These chosen instances would be the elimination of ambiguity for positive sample set. Then we proposed a method to select a negative instance from each negative bag. For each negative bag, the negative bag in each example and selected by the disambiguation of sample set is computed from the distance, choose the most distant examples as negative bag representative instance.Finally, we proposed two feature mapping methods:one for instance-level classification and the other for bag-level classification. These two methods could to embed instance into a feature space that can be solved by the negative SVDD method. And then we proposed two methods:MI-NSVDD_I and MI-NSVDD_B based on feature mapping methods. In MATLAB platform, we use the experimental results show that the two methods can solve the multi-instance learning problem effectively.In summary, this paper based on SVDD method proposed multi-instance learning methods:MI-NSVDD I and MI-NSVDD B are new methods which solution the multi-instance learning problem effectively.