【作者】 徐聪；

【导师】 王邦；

【作者基本信息】 华中科技大学 ， 信息与通信工程， 2021， 硕士

【摘要】 各类互联网公司希望通过精准的信息推送策略,使得产品能够吸引用户兴趣,提高用户粘性。用户面对信息过载问题,期望能直接看到自己真正感兴趣的信息。推荐系统作为应对信息过载问题的有效解决方案,在近些年受到了学业界和工业界的共同关注。本文基于图结构学习的神经网络的推荐算法进行了研究,将推荐问题中的“用户-购买-物品”事件组抽象成二分图中“节点-连边-节点”的形式,在对应的图结构上进行特征学习和推荐。具体来说,包括以下三个方面的内容:(1)提出一种结合图嵌入预训练的图协同过滤算法:该方法的核心思想是使用图嵌入算法对二分图上的节点嵌入向量进行预训练,再通过提出的图协同过滤模型对节点向量表示进行学习。其中,图嵌入算法对节点的预训练能够提取节点的高阶结构特性,用以丰富节点表示;图协同过滤模型利用图的连接性,学习节点的低阶协同信息,得到节点的最终表示向量。实验表明该方法能够有效提升模型对节点的表示能力,提升了模型的推荐性能。(2)提出一种基于群组的图协同过滤模型:相比于过去二分图上的推荐模型,该模型在学习节点向量表示时不仅考虑了节点自身的特征向量,还将节点的群组信息作为节点特征的一部分参与学习。具体来说,首先通过基于随机游走的方式将节点划分到不同的群组,接着将群组信息向量化并作为节点特征的一部分参与图协同过滤模型的训练中;在图协同过滤模型中,通过自注意力机制计算来自不同邻居的信息权重,实现更优的节点表示更新。实验表明,通过该方法能够得到更好的节点表示向量,获得更优的推荐性能。(3)提出一种基于群组的多层感知机协同过滤模型:是对基于群组的图协同过滤模型的改进。该模型在使用群组信息的同时,使用了多层感知机结构来对训练数据进行成对训练,有效解决了计算资源消耗过高的问题;在对节点群组特征的获取上,使用了注意力机制融合来自不同群组的信息;此外,在节点表示向量中加入了节点对的交互特征。实验表明,相比于传统只使用节点本身特征来学习节点表示的方法,该模型能够得到表现更好的表示向量,实现更好的推荐性能。综上所述,本文在二分图的推荐问题上进行了改进与创新,由于二分图结构的普适性,提出的算法模型可以应用于多数推荐场景中,对缓解当前信息过载的问题具有一定的参考价值和积极意义。更多还原

【Abstract】 Internet companies hope to make their products appeal to users’ interests and improve their stickiness through precise information pushing strategies.Meanwhile,users face the problem of information overload and expect to see the information they are really interested in.As an effective solution to the information overload problem,recommender systems have received a lot of attention from both academia and industry in recent years.This thesis has studied the research on neural recommendation based on graph structure learning,abstracting the ”user-purchase-item” event group in the recommendation problem into the form of ”node-edge-node” on bipartite graphs,and performing feature learning and recommendation on the corresponding graph structure.Specifically,the study coverd following three aspects.(1)Propose a graph collaborative filtering algorithm combined with graph embedding based pre-training: the core idea of this scheme is to propose the use of a graph embedding algorithm to pre-train the node embedding vectors on bipartite graphs,and then learn the node vector representation by the proposed graph collaborative filtering model.The graph embedding algorithm is able to extract the higher-order structural properties of the nodes to enrich the node representation;the graph collaborative filtering model uses the connectivity of the graph to learn the low-order collaborative information of the nodes to obtain the final representation vectors of the nodes.Experiments show that the method can effectively improve the representation of nodes and enhance the recommendation performance of the model.(2)Propose a group-based graph collaborative filtering model(GGCF): compared with the previous recommendation models on bipartite graphs,GGCF not only considers the nodes’ self-features when learning node representations,but also involves the nodes’ group information as part of the nodes’ features.Specifically,firstly,nodes are divided into different groups based on random walk,and then the group information is vectorized and participated in the training of graph collaborative filtering model as part of node features;in the graph collaborative filtering model,the information weights from different neighbors are calculated by the self-attention mechanism to achieve better node representation updates.Experiments show that better node representation and better recommendation performance can be obtained by this method.(3)Propose a group based multilayer perceptron collaborative filtering model(GMCF):compared with the GGCF model,the GMCF model uses a multilayer perceptron structure to train the training data pairwise while using the group information,which effectively solves the problem of high computational resource consumption in GGCF;in the acquisition of node group features,the attention mechanism is used to aggregate the information from different group.In addition,the interaction characteristics between nodes are also added to the node features.The experiments show that GMCF can obtain better performance of expression vectors and achieve better recommendation performance than the traditional method of learning node expressions using only the node features.In summary,this thesis improves and innovates on the recommendation problem on bipartite graphs.Due to the universality of the bipartite graph structure,the proposed algorithmic model can be applied to most recommendation scenarios,which has certain reference value and positive significance to alleviate the current information overload problem.更多还原