Research And Implementation Of Deep Learning Based Recommender System

With the advent of the information age,information data is growing rapidly.Information overload has made people difficult to obtain valuable information.As an information filtering method of recommending items to users actively,the recommender system can alleviate information overload effectively and has been widely studied.Recently,deep learning is widely used in recommender system because of its powerful feature extraction ability,which promotes the development of recommender algorithms.However,deep learning based recommender system still has some problems,such as insufficient utilization of cross-feature,difficulty in engineering implementation,which limits the further optimization and application.In order to fully mine the relationship between features and explore the implementation of deep learning based recommender system,the main work of this thesis is as follows:(1)An attentional multi-order feature interaction network(AMFIN)is proposed.AMFIN combines novel multi-order feature interaction mechanism(MFIM)and capacity-differentiated attention mechanism(CDAM)to express cross-feature more accurately.In AMFIN,MFIM improves the mechanism of vector multiplication chain to reduce the computational complexity of high-order feature interaction.CDAM improves the weight normalization method and combines the global information to adapt to the different importance of features between and within samples.(2)A distributed recommender system is designed and implemented.This recommender system is based on Spark and Tensorflow and includes data processing module,algorithm definition and training module,algorithm service module.Offline and online data are processed by the data processing module in parallel,the algorithm definition and training module uses parameter server framework to train models by synchronous update method,and the algorithm service module provides microservices and real-time algorithm prediction service.Finally,experiments are carried out on AMFIN and the distributed recommender system.The results are as follows:(1)On Avazu dataset,AUC and Logloss of AMFIN are 0.7906 and0.3729,while on Criteo dataset,they are 0.8140 and 0.4380 respectively.Compared to other advanced baseline algorithms,AMFIN performs better and is less prone to overfit.(2)The test of the distributed recommender system shows that its function is stable.It has a near linear acceleration effect on the training of deep learning algorithms and the online service based on AMFIN algorithm can meet the real-time requirement in the multi-sample situation.