| Online recommendation algorithms need to be able to recommend products and services more quickly and accurately.In this context,hash recommendation algorithms have emerged.Hash recommendation maps users and items to a low-dimensional vector space,and uses Hamming distance for similarity calculation,achieving fast and real-time recommendation.Although hash recommendation has the advantages of good scalability and fast response,there are still some disadvantages,such as hash conflicts,storage space waste caused by long hash codes,and low recommendation accuracy.In order to improve the performance of hash recommendation without increasing its computational cost,this paper proposes to address the issues and challenges by not sacrificing part of the recommendation accuracy.The main work contents of this paper are as follows:(1)Introducing a variational recommendation algorithm based on differential Hamming distance.The Hamming distance calculates the similarity between user and item hash codes through bitwise operations,but it does not consider the importance of features and ignores the potential differential information in each bit of hash codes.Therefore,the differential Hamming distance is proposed,which assigns bit weights to hash codes by considering their differences.A variational hash recommendation model is designed for the differential Hamming distance,which consists of user hash and item hash components.First,variational encoders are used to generate hash codes for users and items,and Gaussian noise is added to improve the robustness of hash codes.Then,the differential Hamming distance is used to optimize the hash codes of users and items,improving the model’s ability to reconstruct user-item ratings.Experimental results on two public datasets show that the proposed model can achieve higher recommendation performance than state-of-the-art hash recommendation algorithms under the same computation cost.(2)Introducing a distillation hashing recommendation algorithm based on graph convolutional networks.The expressive power of hash codes determines the recommendation performance of hash-based recommendation systems.Graph convolutional networks(GCNs)have shown excellent performance in various fields by learning node embeddings in the user-item interaction graph and aggregating neighborhood information to improve node representation.Therefore,we propose a distilled hash recommendation algorithm based on GCNs.Firstly,to address the problem of information loss during message passing in GCNs,we propose to skip infinite layer message passing and approximate it by constraining the loss limit,thus avoiding the influence of noise and irrelevant information on hash code representation.Secondly,we design a distillation method to extract ranking information obtained in real-valued recommendation into hash recommendation through ranking distillation technology,which utilizes high-precision recommendations in real-valued recommendation to improve the recommendation performance of hash-based recommendation.Experimental results on five datasets demonstrate that the proposed algorithm after ranking distillation achieves better performance than the latest baseline.In summary,the differential Hamming distance and knowledge distillation can effectively address the problems of Hamming distance and insufficient hash code expression ability in hash-based recommendation.The use of the differential Hamming distance method can improve the accuracy of hash code similarity comparison while ensuring recommendation efficiency.The knowledge distillation method can improve the recommendation performance of hash-based recommendation by utilizing the ranking information obtained in real-value recommendation.A series of experiments were designed to verify the effectiveness and feasibility of the model,which has research value in the field of hash-based recommendation. |
