| 引用本文： |
-
任玉媛,马宏,刘树新,李英乐.基于THDAN模型的时序超网络链路预测方法[J].信息安全学报,已采用 [点击复制]
- renyuyuan,mahong,liushuxin,liyingle.Link Prediction Method based on THDAN Model for Sequential Hypernetwork[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 动态网络链路预测将现实系统的内在关联抽象为普通的图结构，旨在通过探索历史的交互规律来预测下一个时刻的演进状态，因其在社会学、生物学、经济和工业等领域的广泛应用而成为了网络科学中的热门话题。目前大多数算法基于低阶网络的演进规律进行分析，主要通过普通图的刻画来建模现实系统的拓扑结构，从而构建时序的特征空间识别未来的链路。然而这些模型粗略地将丰富的多元关系强制性压缩为二元关系，破坏了每条链路表达的信息完整性，同时一些基于统计学的传统方法单纯地将网络的动态变化视为线性的，而忽视了进化过程中非线性的本质特征，导致模型在不同类型的网络中的表现不尽人意。因此本文提出了一种新的深度模型：时序超图双重注意力网络（Temporal Hypergraph Double Attention Network，THDAN）来解决超网络中具有挑战性的时序链路预测任务。该模型首先结合超图卷积层和空间注意力层来编码超网络的高阶数据关系，同时采集节点间不同关系所携带的语义信息实时地更新网络状态。然后利用基于时间注意力机制的门控循环单元（Gated Recurrent Unit：GRU）网络来捕获动态网络非线性的演化特征和不同历史快照对未来网络的贡献程度，二者有效地融合了网络的空间和时间信息，最终联合全连接层来预测未来可能出现的超链路。此外整体采用生成对抗的博弈框架来增强模型生成下一个超网络快照的质量，从而合理地利用超网络的拓扑结构和演化模式来提高链路预测性能。在四个真实数据集上进行了对比实验，所提出的模型在均方根误差和错误率指标下均显著优于基线模型。 |
| 关键词: 时序链路预测 超网络 超图 注意力机制 |
| DOI： |
| 投稿时间：2023-02-16修订日期：2023-05-07 |
| 基金项目:河南省科技重大专项, |
|
| Link Prediction Method based on THDAN Model for Sequential Hypernetwork |
|
renyuyuan1,2,3, mahong1,2,3, liushuxin1,2,3, liyingle1,2,3
|
| (1.People'2.'3.s Liberation Army Strategic Support Force Information Engineering University) |
| Abstract: |
| Dynamic network link prediction abstracts the internal relationship of the real system into a common graph structure, aiming to predict the evolution state of the next moment by exploring the interaction law of history. It has become a hot topic in network science because of its wide application in sociology, biology, economics, industry and other fields. At present, most algorithms are based on the evolution law of low-order networks. They mainly model the topological structure of the real system through the characterization of ordinary graphs, so as to build the feature space of time series to identify the future links. However, these models roughly compress the rich multivariate relationships into binary relationships, destroying the integrity of the information expressed by each link. At the same time, some traditional statistics-based methods simply regard the dynamic changes of the network as linear, while ignoring the non-linear nature of the evolution process, resulting in unsatisfactory performance of the models in different types of networks. Therefore, this paper proposes a new depth model: Temporal Hypergraph Double Attention Network (THDAN) to solve the challenging temporal link prediction task in the hypernetwork. The model first sets the hypergraph convolution layer and the spatial attention layer to encode the higher-order data relationship of the hypernetwork, and simultaneously collects the semantic information carried by different relationships between nodes to update the network status in real time. Then, the gated recurrent Unit (GRU) network based on temporal attention mechanism is used to capture the nonlinear evolution characteristics of the dynamic network and the contribution of different historical snapshots to the future network. The two effectively integrate the spatial and temporal information of the network, and finally combine the full connection layer to predict the possible hyperlinks in the future. In addition, the game framework of generating confrontation is adopted to enhance the quality of the next hypernetwork snapshot generated by the model, so as to reasonably use the topology and evolution mode of the hypernetwork to improve the link prediction performance. Comparative experiments were carried out on four real data sets, and the proposed model was significantly better than the baseline model under the root mean square error and error rate indicators. |
| Key words: sequential link prediction hypernetwork hypergraph attention mechanism |
