基于超图嵌入和有限注意力的社会化推荐

Social Recommendation Based on Hypergraph Embedding and Limited Attention

近年来,社会化推荐作为推荐算法之一被广泛应用于各大平台.由于引入了用户的社交信息,社会化推荐可以较好地缓解数据稀疏问题.然而,大部分社会化推荐难以高效地从原始信息中提取用户的有效信息,导致引入社会信息的同时也会引入大量噪声.为了解决上述问题,本文提出了SRBHL(Social Recommendation Based on Hypergraph embedding and Limited attention)模型,通过超图嵌入模块提取用户的历史行为信息和社交信息,以缓解原始目标用户数据稀疏问题,并结合有限注意力模块来过滤原始信息的噪声,最后将得到的有效好友信息用于推荐.在Yelp-Urbana、Yelp-Phoenix和Epinions3个真实数据集上的实验结果表明SRBHL模型相比其他的推荐算法表现更出色.此外,本文还对SRBHL模型进行了鲁棒性分析,并给出了模型最优参数的取值范围.

In recent years,the social recommendation has been widely used on various platforms as one of the recommendation algorithms.Due to the introduction of social information,the social recommendation could alleviate the problem of data sparseness.Nevertheless,most social recommendations are hard to extract the user′s effective information from the original information efficiently,thus introducing the noise of the original information.In order to tackle the above problems,this paper proposes the SRBHL(Social Recommendation Based on Hypergraph embedding and Limited attention)model,which extracts the user′s historical behavior information and social information through the hypergraph embedding module to alleviate the problem of sparse data of the target user.Furthermore,we combine the limited attention module with filtering the noise of the original information.And finally,recommend using the best influential friends′information.The experimental results on three real datasets of Yelp-Urbana,Yelp-Phoenix and Epinions show that the SRBHL model performs better than other recommendation algorithms.In addition,this paper also conducts the robustness analysis of the SRBHL model and suggests the range of the optimal parameters in the model.