Research And Implementation Of Sequence Data Modeling Based On Deep Learning

The artificial intelligence technology is developing rapidly,then lots of sequence data with internal relevance have been accumulated in various applications of the internet.How to model sequence data effectively for mining the association between sequence data better,and make more accurate and faster prediction for the future is an very significant research work to promote technology.However,there are many problems and challenges in the process of sequence modeling,such as the sparseness of sequence data,the loss of feature information transmission caused by long-range dependence,and the reasoning difficulty of complex relationship.This paper focuses on the above problems in specific areas,and makes some changes and innovations by using the auxiliary information in the traditional algorithm model.At the end,the proposed model can improve the sparsity of sequence data,information conversion and reasoning difficulties is proved.(Ⅰ)Firstly,in order to solve the problem of sparsity of sequence data,a prediction model of sequence data based on composed neural network is proposed.The global recurrent neural network model is used to mine the relationship among users,search statements and ads,which is used to mine the potential preference of users;the local recurrent neural network model is used to mine the relationship between search statements and ads,which is used to transform the query intention into the probability distribution of different advertisements.Finally,the two recurrent neural networks are combined to form a combined recurrent neural network,which is used in advertising recommendation to improve the effect of it.Finally,the proposed model is tested on the real ad Click data set,and compared with the traditional algorithm,the MRR is 0.581 and The performance of Accuracy@20 is up to 90.3%.The superiority of the proposed model in solving the problem of sparse sequence data is proved.(Ⅱ)Secondly,in order to solve the problem of feature information loss and reasoning difficulty caused by long-range transmission of sequence data,a prediction model of sequence generation based on multi-turn attention mechanism is proposed.In visual Q&A and target detection,single turn attention mechanism model is generally used,but it can not work well in information transmission.Therefore,the network model of multi-turn attention network is proposed,which encodes and stores the visual image information and text information respectively,and projects the required questions to two kinds of networks at the same time to retrieve the fact basis,so that the cross modal information can interact in multiple rounds,so as to solve the problem of information transmission and reasoning in sequence modeling.Finally,the proposed method is tested in the real visual conversation data set.Comparing this model with the traditional algorithm,the indicators of Multi-QIH-1 surpass other comparison algorithms,which proves the effectiveness of the cross-modal attention mechanism model,about Accuracy@10,Multi-QIH-1 is 1.55%higher than Multi-QIH-2,which proves the proposed multi-turn attention mechanism modeling is effective for information transmission.