Deep Learning Recommendation Methods And Applications

With rapid development of cloud computing,big data and artificial intelligence,adoptive learning became a research hotspot in educational information technology area.At present,in the process of education development,there are contradictions between large-scale education and the differentiated growth of talents.Thus,it is important to build a technical system that can not only gather vast high-quality resources but also provide personalized services.This is a scientific and technical problem that must be confronted to achieve the goal of fair and quality education.This paper set the online education scenario that have large number of users and massive teaching contents as research background,aiming to provide personalized and intelligent learning services to learners.In order to tackle the problems in traditional recommendation system including data sparseness and imbalance,lacking expansibility,low efficiency and low availability,poor robustness,this paper studies the methods of personalized resource recommendation in large-scale online learning environment providing personalized and intelligent service for learners.The main research content and innovation of this paper are mainly reflected in four aspects.For solving the data sparseness and imbalance problem,a new recommendation method,deep matrix factorization(DMF),is proposed.In deep matrix factorization model,two feature transforming functions are built to directly generate latent factors of users and items from various input information.As for the implicit feedback that is commonly used as input of recommendation algorithms,implicit feedback embedding(IFE)is proposed.IFE converts the high-dimensional and sparse implicit feedback information into a low-dimensional real-valued vector retaining primary features.Using IFE could reduce the scale of model parameters conspicuously and increase model training efficiency.In order to alleviate data sparseness and inducible over-fitting problem,deep variational matrix factorization model(DVMF)is proposed.In deep variational matrix factorization model which is a deep learning based fully Bayesian treatment recommendation framework,the latent factors of users and items are set as fully random variables to match the mutability of user features.In addition,the variational inference technique and the reparameterization tricks are introduced to make DVMF possible to be optimized by the stochastic gradient-based methods.For alleviating the poor robustness of ordinary learning algorithm about adversarial attack,adversarial deep latent factor model(ADLFM)is proposed.Adversarial deep latent factor model was proposed with two variants,namely,ADLFM with input adversarial attacks(IAAs)and ADLFM with latent factor adversarial attacks(LAAs).In the ADLFMs,two types of adversarial training frameworks,including modeling and optimization procedures,are applied to the proposed deep learning recommendation model against identified two designed adversarial attacks,IAAs and LAAs.For solving the data sparseness and imbalance problem,separated graph neural recommendation(SGNR)model is proposed based on graph information convergence.SGNR models the group characteristics of users and items and extracts the collaborative characteristics of the groups to effectively alleviate the problem of data sparsity and unbalance.Beside,graph separation process separates heterogeneous networks in recommendation problems to two homogeneous networks(user similarity network and item similarity network);SGNR decoupled the feature aggregation and nonlinear mapping processes in traditional graph neural networks boosting model availability and efficiency.The algorithm verification and analysis of the four proposed recommendation models are carried out by using the data collected in the online learning environment and the classical public data sets in recommendation systems.Experimental results show that the proposed algorithm is generally superior to the current mainstream recommendation methods in terms of quantitative assessments.