Application Of Deep Learning In Data Mining

Due to the rapid progress of internet technology,substantial time-series data has been created in lots of social fields.Generally speaking,important information and variation trends are always hidden in the mass data.However,there exists many difficulties in accurate hidden-information obtaining and variation trends forecast of time-series data,and the accuracy of time-series data predictions still needed to be improved.With the breakthrough of deep learning algorithm,artificial intelligence has developed rapidly.Compared with the mathematical models in Statistics and Metrology,deep learning algorithm has gained more competence in data mining of time-series data.Firstly,this paper investigates whether the machine learning is applicable in data mining of time-series data by testing the Linear Regression and Support Vector Machine.The theoretical analysis and simulations reveal that the role of machine learning in data mining of time-series data is limited.Secondly,the paper turns to the study of deep learning algorithm as well as the Feedforward Depth Network structure,activation function and training optimization method.By deducing the signal forward propagation and updating the weights of deep neural network based on back propagation algorithm,it shows that,theoretically,Feedforward Depth Network does not possess the correlation in time series and is not applicable for time-series data forecast,while Recurrent Neural Network which has the time-series correlation does.To address the problems of vanishing gradient and exploding gradient which exist in the training process of Recurrent Neural Network,the paper adopts the improved Long Short-Term Memory Neural Network.By further studying the internal structure and weight update process,the paper has analysis the advantages and limitations of LSTM neural network's application on time-series data.Optimize network structure and the Ada Split optimize learning rate algorithm have been adopted to solve the problem that LSTM neural network relies too much on learning rate.As a result,the Optimized LSTM neural network is capable to do accurate time-series predictions.This paper has tested the applicability to time-series data of Support Vector Machine,Recurrent Neural Network and optimized Long Short-Term Memory Neural Network by lots of simulation experiments.Visual simulation results shows that the optimized Long Short-Term Memory Neural Network gains higher forecast accuracy and better imitative effect.