The Research And Application Of Classified Forecasting Approach Based On Deep Learning

Deep learning is a branch of artificial intelligence,and the new direction of data mining.Deep learning is based on artificial neural network,deep learning has much more hidden layers than shallow neural network,which is the origin of the word "deep".Unlike the traditional neural network,deep learning adds unsupervised learning into neural network.Deep learning has great research space in application of classification and prediction.The purpose of this paper is to research the application of classified forecasting approach based on deep learning,the specific application is to use the deep learning algorithm to predict the possibility of students' ability of writing the excellent graduation thesis through the students'information before dissertation proposal.According to the study of this application,by the time of dissertation proposal,the school and tutors can dig out the students who have the potential to write excellent graduation thesis,and then use the next time to cultivate these students.This research and application of classified forecasting can provide the reference for how deep learning which is good at processing big data and high-dimensional data play feasibility and superiority when facing dataset that has defects.This paper adopt the deep learning means of the stacked denoising autoencoders to do the application.Stacked denoising autoencoders organize many denoising autoencoders to form the hidden layers of the deep learning model,and use softmax regression classifier as output layer of the deep learning model.The key content of this deep learning algorithm lies in firstly doing unsupervised layer-wise pretraing which is denoising,then on the basis of unsupervised learning,deep learning do the supervised learning.The input samples of deep learning application come from the students' information database.In face of the situation that students' information is too few and positive and negative samples are not balanced,this paper refer to the advantages and disadvantages of batch learning and online learning and make use of redundant samples to increase the proportion of minority samples and solve the problem that samples' size is tinsufficient.This paper also adopt the means of attributes repeated artificially to solve the problem of samples' dimension are too few.the predicting result of this paper can easily reach the accuracy of more than 90%.This paper give the concrete prediction result under different setting of different parameters,according to the comparison of these results,these results can clearly illustrate the impact of different parameters on the prediction results.This paper also gives the different prediction results under the conditions of redundancy and dimension extension to illustrate that big data and high dimension data can bring greater exploration space for deep learning.This paper use students' data set and the UCI's data set which don't need to have the process of data pretreatment as the input data set of deep learning.According to the comparison experiment,this paper proves that deep learning algorithm has superiority over the pure unsupervised learning algorithm,the pure supervised learning algorithm,the common machine learning algorithm and the ordinary neural network algorithm.This paper also proves that deep learning algorithm is an empirical model.With higher dimension of data set and more samples,deep learning algorithm can adjust more parameters and have more advantages.The application of deep learning in this paper has deficiencies,deficiencies are mainly from two aspects which are the lack of samples and the disadvantage of algorithm.The input samples have two serious problems.One problem is the sample size is insufficient,the other problem is data normalization is difficult.The disadvantage of deep learning algorithm is that there is no global optimum in deep learning algorithm.This paper proves that even the dataset has problems like samples lacked,deep learning can still play superiority in classification and forecasting.This paper proves that although deep learning algorithm is the empirical model,it still needs the theoretical support in the application.