Research On Complicated Semi-Supervised Learning Setting

Semi-supervised learning aims to improve the performance of learning algorithms by utilizing a large amount of unlabeled data.It is a basic research problem in the field of machine learning.After nearly 20 years of research,semi-supervised learning has achieved rather good performance in a relatively simple environment.However,in recent years,machine learning has been faced with complex environments in real-life applications.Semi-supervised learning is no exception,and it is necessary to provide corresponding so-lutions in the face of complex environments.In view of the general trend of complex environment,this paper studies the semi-supervised learning in complex environments from two aspects.Firstly,the learning objective of a real task is usually related to multiple categories,for example,an image can be related to multiple categories.At this point,the lack of category tags is much more complicated than the binary semi-supervised learning in a simple environment.To this end,this paper first carries out the semi-supervised weak-label learning problem.In the semi-supervised weak-label learning problem,the labels of the data may be all missing or partially missing.In this paper,the semi-supervised weak-label learning algorithm SSWL is proposed,and the data similarity and label similarity are used to supplement the missing mark,and the integration of multiple models is used to improve the robustness of the learning effect.The learning problem is modeled as a biconvex optimization(Bi-Convex)form and an efficient block alternation optimization algorithm is given.A large number of experimental results verify the effectiveness of the algorithm.Secondly,real-world tasks usually face online dynamic scenarios,such as adaptive recommendation tasks.The recommendation algorithms need to change dynamically,and at the same time,the labels of the data is often missing.To this end,this paper further studies the theoretical basis of the dynamic online semi-supervised learning algorithm.This paper presents a new online semi-supervised learning strategy for dynamic environments.This strategy effectively utilizes useful historical experience information through a new loss function under a smoothing assumption.In this paper,the regret bound of dynamic online semi-supervised learning,which can be regarded as the generalization performance in the dynamic environment,is given based on the iterative least squares algorithm,and sufficient conditions are given to judge the selection of suitable unlabeled data.To the best of our knowledge,this is the first work to study the impact of unlabeled data on dynamic regret boundaries.