Research On Multi-Task One-Class Learning Based On Dictionary Learning

One-class learning is based on one class of examples.It has received more and more attention in the field of machine learning and data mining.One-class classifier is constructed based on one class of examples.Here,one-class learning takes one class in the dataset as the target class,and the other classes in the dataset as the non-target classes,and then learns a classifier from the target class to predict whether a new example belongs to the target class or non-target class.Today,one-class learning has been widely used in intrusion detection,image retrieval,text retrieval,and remote sensing.Therefore,in-depth study of one-class learning is extremely important.With the rise of commercial data,more and more datasets have similar distribution.Multi-task learning has received more and more attention and is suitable for processing datasets with similar distributions.At the same time,in the case of multiple tasks,the amount of labeled data for each source task may be much larger than that for the target task.However,most existing one-class models are suitable for single-task problem,and are not suitable for multi-task one-class problem.Therefore,single-task one-class learning needs to be studied in depth,and the multi-task one-class learning needs to be extended.In this paper,we propose multi-task one-class learning based on dictionary learning?MTD-OC?.This method adds dictionary learning to one-class learning.First,we give each task a dictionary to ensure that the dictionaries for different tasks are independent and as distinguishable as possible.Dictionary learning contains three items,l_2,1-norm constraint,dictionary incoherence term and projection matrix extraction term,which aim to improve the performance of presentation,promote presentation incoherence among tasks and improve coding efficiency.The one-class classifier of the target task is then constructed with the help of the transfer knowledge from multiple source tasks.At the same time,one-class learning improves the performance of dictionary learning,and dictionary learning improves the performance of one-class learning.In the multi-task one-class learning based on dictionary learning,the optimization function optimizes both one-class learning and dictionary learning.Therefore,we propose an iterative framework to solve the optimization function to obtain the classifier of the target class.In the experiment,three actual text datasets are first segmented to generate multiple sub-datasets,and the distributions of these sub-datasets are similar.By comparing other one-class methods,it is shown that the multi-task one-class learning based on dictionary learning can construct a classifier and improve the accuracy by learning the dictionary of each task.