Research On Recommendation Algorithm Based On Adversarial Training

Robustness and data sparsity have become two hot topics in the research of recommender system.Robust recommendation aims at capturing true preference of users from noisy data to provide accurate and stable personalized recommendation.Data sparsity refers to the fact that compared to the massive users and items in the recommender system,the items that each user has interacted with are only a small part of the whole set of items,which prevents the existing recommendation methods that depend on the user's historical feedbacks from accurately capture personalized preference of users.Although the existing works have studied the problems of robustness and data sparsity of recommender system,however,these two problems are still far from being well solved.Robust recommendation is far from being well solved partly due to two challenges.The first is the Personalized Noise Reduction.In real world,due to different behaviors and item usage patterns,the noise in data of different users and items has different distribution,which requires that robust recommendation should provide personalized noise reduction with adaptability to different noise distributions.However,many existing methods cannot capture the difference in noise distribution.The second one is the Multi-Modality of Preference,which requires that recommendation model should be expressive enough to approximate multiple distributions of users' preferences,while the existing VAE-based methods only learn single modality of latent space.In this paper,to overcome the defects of the existing methods for robust recommendation,we propose a novel Dual Adversarial Variational Embedding(DAVE)model,which is able to provide the personalized noise reduction and capture the multi-modality of the preference distributions,by combining the advantages of VAE and adversarial training.The extensive experiments conducted on real datasets verify the recommending performance and robustness of DAVE.In the context where only user-item interaction data is available,few data augmentationbased recommendation methods have been proposed to alleviate the data sparsity problem for accurate recommendation.However,the existing methods still need to solve the two challenges.The first is the Adaptive Data Augmentation,which requires that the data augmentation should be adaptive to the need of accurate personalized preference learning of different users.This is because the user's personalized preferences make the data sparsity of different users have different effects on their personalized preference learning.However,the existing methods often conduct data augmentation for all users,reducing the effectiveness of data augmentation.The second is the End-to-End Trainability.The existing methods separate the process of learning data augmentation model from the process of learning recommendation model,which is likely to augment inappropriate data that limits the generalization performance of recommendation model,and also hinders the adaptability of data augmentation.This inspires us that the learning of the data augmentation model should be incorporated with the learning of the recommendation model in an end-to-end manner for effective and adaptive data augmentation.In this paper,to address the above challenges,we propose a novel model called Gated Rating Augmentation with Adversarial Generation for Recommendation(GRAAG).The main idea of GRAAG is to adaptively augment rating with a designed gated mechanism and the combination of the adversarial training and end-to-end training,so that the rating augmentation can adaptive to the need of accurate personalized preference learning.The extensive experiments conducted on real datasets verify the recommendation performance and the adaptability of personalized dealing with data sparsity of GRAAG.