Research And Application Of Density-based Semi-supervised Online Sequential Extreme Learning

In reality,there is a large amount of unlabeled data,and marking unmarked data often requires a lot of manpower and material resources.Semi-Supervised Learning(SSL)is a learning system that uses a small amount of tagged data for pre-training and a large amount of unlabeled data for assisted training.How to improve the accuracy of semi-supervised learning has been studied by countless scholars in the past decade.Most of the current methods are to deal with unlabeled samples in different ways,and increase the constraints to improve the optimization effect.In the actual data,the training data may arrive one-by-one or block-by-block.For such data,the paper adds the online sequence extreme learning machine,so that in each iteration,it is not necessary to retrain all the data,only need to use the new data to fine tune the network,which reduces the training time of the data.Moreover,most of the real data in reality are complex data,and the distribution of many data does not conform to the Gaussian distribution,and may be a non-spherical distribution or a bimodal distribution.Some traditional classification algorithms and clustering algorithms are difficult to achieve good results.Therefore,this paper proposes a density-based semi-supervised online sequence extreme learning machine(D-SOS-ELM).The proposed method can realize online learning of one-by-one or block-by-block unlabeled data.In addition,the use of local density and relative distance can also effectively reflect the relationship between data.Compared with the traditional method of measuring the confidence between data based on distance,the proposed strategy improves the ability to process complex data.The work of this paper is mainly divided into two parts:1.Propose a strategy to use local density and relative distance to measure the similarity between data,select high-reliability data for online learning,which can effectively improve the accuracy of learning and the speed of learning.The proposed method achieves efficient learning of unlabeled data by continuously selecting high confidence unlabeled data.The experimental results show that the proposed method can inherit the Clustering by fast search and find of density peaks(CFSFDP)algorithm and has a good discriminating effect on the irregular data.2.A number of experimental comparisons of several standard benchmark datasets confirm that the proposed D-SOS-ELM model has better accuracy and is superior to the existing advanced methods.Further experiments on the MNIST data set also yielded good learning results.