Research And Application Of One-class Classification Based On Transfer Learning

In the era of rapid development of artificial intelligence,the demand for data is getting higher and higher,such as independent distribution and sufficient data,but it is not possible to obtain the desired data in all situations,as if we are not always with its own knowledge,transfer learning,like knowledge transfer,gradually enters people's field of vision and is widely used in traditional classifiers such as support vector machines.In the traditional supervised learning model,for example,support vector machines require multiple types of samples as the training set,but in practical applications,it is difficult to achieve or to pay a high price for this,such as enemy identification,equipment failure,etc.Only one sample of the tag can be obtained.And in our life there is such a problem,we only need one class,the rest are outliers,and we don't care what the label of outliers is.This is a single classification problem.The content of this paper is how to apply the migration learning algorithm to the single classification problem,and experiment to judge whether the nature of the fusion is good.This dissertation adopts a sample-based direct transfer learning algorithm based on kernel average matching algorithm to perform transfer learning,and applies support vector data description algorithm for single classification.In the experiment,we used Pharmacopoeia data obtained by the FDA and TcmSP and vectorized the drug molecules with Dragon software.We analyzed the properties of traditional Chinese medicines with the molded western medicines through the single classification problem of migration learning,and analyzed the ROC curves through the existing data.And the corresponding AUC value to determine the migration effect,the final experiment shows that the fusion performance of the two is better.