Research On Recommendation Algorithm With Generator-Discriminator Ensemble

Personalized recommendation is an instrumental approach to address information explosion,attracting more and more attention and research in the academic community at the same time to benefit commercial platforms.Recommendation system(RSs)can be viewed as query-rank procession,Generative adversarial networks(GANs)can dynamically simulate user-item interactive process by it's characteristic of complementary principle between generator and discriminator.However,original GANs are limited to generate continuous data while discrete text is main data form in recommendation areas.Policy gradient in Reinforcement Learning(RL)is generally adopted to acquire the approximate gradient of generator,which produces high variance,leading to sample perturbation and making it hard for the model to convergence.Furthermore,it's hard to train GANs,“model collapse” and poor generalization still exist.At inference time,this paper proposed a heterogeneous Generator-Discriminator ensemble recommendation algorithm(HGDE)to simulate user-item interactive process dynamically,making following improvements on the basis of current relevant mainstream algorithms: i)using dynamic negative sampling to preliminarily screen out purpose-built candidate set and objective function optimization,to the extent that improves the efficiency of adversarial model;ii)enriching generative model and discriminative model with a variety of Deep Learning(DL)models to mine non-linear relationship between user-item thus improving prediction accuracy;iii)employing Ensemble Learning to reduce the high variance caused by policy gradient,avoid overfitting and improve generalization simultaneously.HGDE has demonstrated significant performance gains over cutting-edge algorithms about GANs in application of recommenders on benchmark efficiency?Precision@5 and NDCG@5.On recommendation dataset Movielen-100 K,HGDE improves approximate 16.64%?9.17% and 33.34% respectively while 26.06% ?24.30% and 68.73% on FilmTrust.In conclusion,HGDE has better generalization and expansibility besides improving recommendation prediction accuracy.