用于社交推荐的增强影响扩散模型

An Enhanced Influence Diffusion Model for Social Recommendation

随着网络上创建连接、协作、共享的全新变革方式的出现,互联网上丰富的社交行为现象引起了研究者和实践者的关注.近年来,随着社交网络平台的普及与推广,基于社交网络的推荐系统也成为了个性化推荐领域的研究热点之一,社交推荐系统可以利用社交网络来缓解传统的推荐算法中数据稀疏性问题.在社交网络中,社交关系影响起着重要作用,而用户信任是社交关系形成的基础,每一个用户会受到其信任的用户影响,这些被信任的用户也会被自己的社交关系所影响,这就表明了联系在一起的用户会相互影响,导致社交联系之间的用户偏好具有相似性.用户的信任关系影响着用户偏好的推断,同时用户受到其信任用户的社交关系影响,而这些社交关系影响在社交网络中递归传播和扩散.因此,基于社交推荐算法研究的关键就在于信任信息的挖掘和利用.在基于社交网络的推荐领域中,比较有代表性的模型为DiffNet,该模型未充分考虑到信任问题,同时,在递归计算长距离的社交关系时,有额外的噪声,影响推荐预测的质量.本文提出了基于DiffNet改进的社交推荐模型-EIDNet.首先,该模型在模拟社交关系影响扩散过程时,通过用户对物品的历史交互记录建立用户间的信任关系,并融合到递归社交动态建模中,以获得不同用户面向不同物品的不同信任关系.其次,在计算社交影响力时,提出增加残差连接的方式,减少长距离社交关系产生的噪声.同时,为了解决同阶领域权重分配问题,提出使用注意力机制,学习用户朋友关系向量,针对不同的用户计算不同的权重,适应地衡量用户朋友间的社交影响力.最终,将三个部分融合在一个统一的框架中,以相互增强,并建立一个更好的扩展模型.最后,结合用户与物品的历史交互行为和社交关系来预测用户未来的行为和偏好.本文的主要贡献包括:(1)将基于用户与物品的历史交互记录的信任融合到递归社交动态建模中,以解决不同用户因信任关系而面向不同物品进行推荐的问题;(2)在递归计算长距离的社交关系时,提出残差连接的方式来减少噪声影响;(3)引入了注意力机制,为社交网络中的用户给予不同的重要性,以解决同阶领域权重分配问题.实验结果表明,EIDNet与DiffNet最好的性能结果相比,在Yelp上的表现提升了10.61%,在Flickr上的表现提升了24.98%,结果证实了本文提出的改进模型可以提升社交推荐模型的推荐性能.

With the emergence of new and transformative ways of creating connection,collaboration,and sharing on the Internet,the phenomenon of rich social behavior on the Internet has attracted the attention of researchers and practitioners.In recent years,with the popularity and promotion of social network platforms,recommendation systems based on social networks have also become one of the research hotspots in the field of personalized recommendation.Social recommendation systems use social networks to alleviate the data sparsity problem in traditional recommendation algorithms.In social networking,social relationships have played an important role,and user trust is the foundation of social relationship.Every user is affected by the users they trust.These trusted users will be affected by their own social relations,which means that users associated with each and there is preference similarity among socially connected users.The user's trust relationship affects the user's preference,and the user is affected by the social rela tionship between its trust users,and these social relationships are increasing and spreading in so cial networks.Therefore,the focus of social recommendation algorithm researchlies in the min ing and utilization of trust informationin the field of recommendation based on social network,a representative model is DiffiNet,which has not fully considered the problem of trust,and at the same time,there is additional noise when recursive long-distance social relationships,affecting the quality of the recommended forecasts.Therefore,a DIFFNET improved social recommenda tion model-EIDNetis proposed in this paper.Firstly,when simulating the diffusion process of so cial relationship influence,the trust relationship between usersisestablished by the historical inter action records of users with items,and integrates it into the recursive social dynamic modeling to obtain different trust relationships for different users for different items.Second,Secondly,when calculating social influence,a method of increasing residual connections is proposed to reduce the noise generated by long-distance social relationships.At the same time,in order to solve the problem of weight distribution in the same order domain,an attention mechanism is proposed to learn the user friend relationship vector,calculate different weights for different users,and adaptively measure the social influence among users’friends.Ultimately,the three parts are fused into a unified framework to reinforce each other and build a better scaling model.Finally,the user's future behavior and preferences are predicted by com'bining the user's historical inter action behavior and social relationship with the item.The main contributions of this paper in clude:(1)Integrate trust based on historical interaction records between users and items into re cursive social dynamic modeling to solve the problem of different users recommending different items due to trust relationships;(2)In the recursive calculation of long distance social relations,a residual connection method is proposed to reduce the influence of noise;(3)Attention mecha nism is introduced tomatch different importance to users in the social network to solve the prob lem of weight distribution in the same order domain.The experimental results show that,com pared with the best performance results of DiffNet,the performance of EIDNet on Yelp is im proved by 10.61%,and the performance on Flickr is improved by 24.98%.The results confirm that the improved model proposed in this paper improves the recommendation performance of the social recommendation model.