Research And Application Of Semi-Supervised Learning Under Small-Scale Data Sets

In practical applications,the labeled data needed to train intelligent models often have a high acquisition cost.on the other hand,the unlabeled data in real scenes can be easily obtained.Therefore,semi-supervised learning using only a small amount of labeled data and unlabeled data at the same time has great practical significance.Compared with the deep learning semi-supervised method which needs a lot of training data and focuses on dealing with complex data forms,the semi-supervised learning method under small-scale data set has value in many fields because of its fast training speed,low labeling demand and wide application range.For example,the evaluation question database of many online test platforms has a preliminary division based on chapters or knowledge points.However,with its rapid development in recent years,the number of questions under many chapters has reached hundreds of scales.The large question database has gradually made the teachers and students exhausted and need to be further subdivided to improve the user experience of the platform.In the field of intelligent agriculture under the Internet of things,due to the low dimension of data collected by the system and easy to analysis,the semi-supervised learning method under smallscale data sets can replace the current common intelligent decision-making mode of calling the expert resource database directly with less manpower to label the data,and make adaptation to local conditions.According to the characteristics of insufficient labeled data in semi-supervised scene,this paper dynamically optimizes the training data set by adjusting the construction of training set,sample discrimination of unlabeled data and integrating active learning strategy,so as to make it better used for model training,and achieve better semi-supervised learning performance.At the same time,this paper also applies the method to the field of online education and intelligent agriculture.The main contributions of this paper are as follows: 1)based on labeled training data and unlabeled data set in small-scale data scene,a semi-supervised training method DDC based on dynamic training set construction is proposed,which uses unlabeled data information to optimize the training data set and improve the semi-supervised learning performance;2)Aiming at the classification of unlabeled data samples,a semi-supervised training method SD-DDC with sample discrimination is proposed on the basis of DDC,which is applied to the field of intelligent agriculture to realize the intelligent decision-making of intelligent irrigation technology according to local conditions;3)Aiming at the interactive advantages in the field of online education,based on SD-DDC,a semi-supervised training method ASD-DDC combined with active learning strategy is proposed and applied to the field of online education to realize the automatic division process of topic subclasses in the question database of online test platform with the participation of a small amount of expert manpower.