Contributions To Several Issues Of Recommender System For Large Scale And Complex Data

Recommender system has been widely used in people's daily life,and has achieved fruit-ful results in e-commerce,social network,computational advertising,intelligent medicine,smart city and other fields.Recommender system aims to quickly discover a small amount of latent useful information for users from a large amount of information,so that users can avoid getting into the problem of information overload.Recommender system has been widely concerned by academia and industry for many years.However,with the rapid development of the Internet and the growing popularity of intelligent terminals,data in terms of quantity,dimension,form,structure and other aspects has generated a series of challenges,such as massive data,high di-mension of features,multimodal data,and complex structure.Along with these challenges are the problems like storage cost of recommendation model,online recommendation calculation cost,high-dimensional noise data preprocessing,heterogeneous modal data recommendation,information extraction from complex structure and so on.Therefore,this thesis intends to ex-plore solutions to the above problems.The main contributions of the thesis are summarized as follows:1)Firstly,a novel discrete factorization machine is proposed.By constructing the quadratic upper bound of the cross entropy objective function and introducing the discrete constraint,the optimization problem of the latent feature matrix is transformed into a series of binary quadratic programming problems,and the hash coding representation is learned by using the semi-definite relaxation optimization method.The experimental results show that com-pared with the real value factorization machine,the proposed model can achieve similar recommendation performance while significantly reducing model storage cost and acceler-ating recommendation calculation.Therefore,the proposed model can effectively deal with the model storage problem and online recommendation calculation problem caused by the massive data and high dimensionality of features2)Secondly,a novel sparse principal component data analysis method is proposed.By in-troducing the sparse regular operator into the objective function,a stable alternating direc-tion multiplier method is proposed to update multiple principal components on the Stiefel manifold,so that the learned principal components can maintain the orthogonal property while introducing the sparsity.At the same time,a two-stage sparse principal component method is proposed to consider the influence weights between principal components in fea-ture selection.The experimental results show that compared with the existing methods,the proposed method has a significant improvement in the quality of extracted principal com-ponents,feature selection and computational efficiency.Therefore,the proposed method can effectively deal with the problem of data dimensionality reduction and noise reduction caused by the massive data and high dimensionality of features3)Thirdly,a relation-wise cross modal image and text retrieval recommendation model is proposed.Firstly,region feature vectors and word feature vectors are extracted for image and text data respectively.Then we use a visual-semantic relation network to learn the latent alignments between image region and word.Finally,we use a two-way attention network to infer the matching similarity between image and text.The experimental results show that compared with the existing model,the model can better capture the abstract concepts in visual-semantic alignment,and significantly improve the task of cross modal data retrieval recommendation.Therefore,the proposed model can effectively deal with the problem of heterogeneous modal data caused by the multimodal data4)Fourthly,a graph neural network based recommendation model is proposed.The model first learns the latent embedding vector for each user and item through the designed graph embedding message propagation layer,and then uses a multilayer neural network to build the interaction model for the latent embedding vectors of user and item.The experimental results show that compared with the existing model,the proposed model has a significant improvement in recommendation performance,and verify the rationality of the model de-sign.This model can effectively mine the latent collaborative information of complex data and improve the performance of the recommendation model by directly introducing the ge-ometric structure information of users and items into the latent embedded vector learning.