IJCAI2020丨近期必读七篇【GNN】论文

导语：国际人工智能联合会议（International Joint Conference on Artificial Intelligence, 简称为 IJCAI）是人工智能领域中最主要的学术会议之一，原为单数年召开，自2016年起改为每年召开。因疫情的影响, IJCAI 2020将于2021年1月5日-10日在日本举行。

根据AMiner-IJCAI 2020词云图，小脉发现表征学习、图神经网络、深度强化学习、深度神经网络等都是今年比较火的Topic，受到了很多人的关注。昨天向大家分享了八篇表征学习相关论文，今天小脉为大家奉上IJCAI 2020七篇必读的图神经网络（Graph Neural Network）相关论文。

查看更多论文：https://www.aminer.cn/conf/ijcai2020/papers

1. 论文名称：Bilinear Graph Neural Network with Neighbor Interactions

论文链接：https://www.aminer.cn/pub/5ef96b048806af6ef2772079?conf=ijcai2020

作者：Hongmin Zhu、Fuli Feng、Xiangnan He、Xiang Wang、Yan Li、Kai Zheng、Yongdong Zhang

简介：

• The authors compare against the strong baselines mainly in two categories: network embedding and GNN.
• For GNNs, the authors select GCN [Kipf and Welling, 2017], GAT [Velickovicet al., 2018] and Graph Isomorphism Network (GIN) [Xu et al, 2019b].
• For each BGNN, the authors compare two variants with different scopes of the bilinear interactions: 1) BGCN-A and BGAT-A which consider all nodes within the k-hop neighbourhood, including the target node in the bilinear interaction.
• The authors closely follow the GCN work [Kipf and Welling, 2017] to set the hyper-parameters of SemiEmb, DeepWalk, and Planetoid.
• All BGNN-based models are trained for 2,000 epochs with an early stopping strategy based on both convergence behavior and accuracy of the validation set

2. 论文名称：GraphFlow: Exploiting Conversation Flow with Graph Neural Networks for Conversational Machine Comprehension

论文链接：https://www.aminer.cn/pub/5d4409be3a55acdd76472372?conf=ijcai2020

作者：Chen Yu、Wu Lingfei、Zaki Mohammed J.

简介：

• The authors compare the approach with the following baseline methods: PGNet (See et al, 2017), DrQA (Chen et al, 2017), DrQA+PGNet (Reddy et al, 2018), BiDAF++ (Yatskar, 2018), FLOWQA (Huang et al, 2018) and SDNet (Zhu et al, 2018).
• The authors will discuss the details of these methods in Section 4.1

3. 论文名称：Improving Attention Mechanism in Graph Neural Networks via Cardinality Preservation

论文链接：https://www.aminer.cn/pub/5d1f1fd13a55ac72f611c7f0?conf=ijcai2020

作者：Zhang Shuo、Xie Lei

简介：

• The authors focus on the following questions:
  Q3: Since attention-based GNNs (e.g. GAT) are originally proposed for local-level tasks like node classification, will those models fail or not meet the upper bound of discriminative power when solving certain node classification tasks?
  Q4: For global-level tasks like graph classification, how well can the original attention-based GNNs perform?Q5: How the attention-based GNNs with the CPA models perform compared to baselines?.
• To answer Question 4 and 5, the authors perform experiments on graph classification benchmarks and evaluate the performance of attention-based GNNs with CPA models

4. 论文名称：Multi-Channel Graph Neural Networks

论文链接：https://www.aminer.cn/pub/5ef96b048806af6ef277208b?conf=ijcai2020

作者：Kaixiong Zhou、Qingquan Song、Xiao Huang、Daochen Zha、Na Zou、Xia Hu

简介：

• As shown in Figure 1, the authors first describe how MuchGNN learns the various nodes’ characteristics in the single channel at layer 0 by defining the graph convolutional filters.
• The authors describe how MuchGNN operates the graph convolutions to pass message among the multiple channels at layer 1.
• The graph convolutions stage applies GNN model to obtain node embeddings via K steps of message passing.
• Channel expansion Tl ture roles of nodes with blue and grey colors, which bridge the different communities
• These channels encodes the inherently various characteristics of input graph, which makes the higher layers of MuchGNN being more informative for classifying graphs.
• It would be more easier for the downstream classifier to distinguish heterogeneous graphs

5. 论文名称：Coloring graph neural networks for node disambiguation

论文链接：https://www.aminer.cn/pub/5df371de3a55acfd20674ba8?conf=ijcai2020

作者：Dasoulas George、Santos Ludovic Dos、Scaman Kevin、Virmaux Aladin

简介：

• The authors show empirically the practical efficiency of CLIP and its relaxation.
• The authors run two sets of experiments to compare CLIP w.r.t. state-of-the-art methods in supervised learning settings: i) on 5 real-world graph classification datasets and ii) on 4 synthetic datasets to distinguish structural graph properties and isomorphism
• Both experiments follow the same experimental protocol as described in Xu et al (2019): 10-fold cross validation with grid search hyper-parameter optimization.
• As the same experimental protocol as that of Xu et al (2019) was used, the authors present their reported results on Table 1

6. 论文名称：KGNN: Knowledge Graph Neural Network for Drug-Drug Interaction Prediction

论文链接：https://www.aminer.cn/pub/5ef96b048806af6ef2772133?conf=ijcai2020

作者：Xuan Lin、Zhe Quan、Zhi-Jie Wang、Tengfei Ma、Xiangxiang Zeng

简介：

• The authors propose a novel model, called KGNN, for drug-drug interaction prediction.
• KGNN extends spatial-based GNN approaches to the knowledge graph by multiple aggregating neighborhood information selectively, which is able to learn both topological structure information and semantic relation of knowledge graph, as well as the neighborhood of drug and related entities.
• The authors implement the proposed method and conduct experimental comparisons on two widely used datasets.
• The experimental results show that KGNN outperforms the classic and state-of-the-art DDI prediction models

7. 论文名称：Smart Contract Vulnerability Detection using Graph Neural Network

论文链接：https://www.aminer.cn/pub/5ef96b048806af6ef2772179?conf=ijcai2020

作者：Yuan Zhuang、Zhenguang Liu、Peng Qian、Qi Liu、Xiang Wang、Qinming He

简介：

• Blockchain technology is developing rapidly due to its decentralization and tamper-free nature [Tsankov et al, 2018]
• We extend GCN to a degree-free GCN (DR-GCN) to handle the normalized graphs
• Our key contributions are: i) We introduce a novel temporal message propagation network (TMP) and a degree-free GCN (DR-GCN) to automatically detect smart contract vulnerabilities. ii) We propose to characterize the contract function source code as contact graphs, and explicitly normalize the graph for highlighting the key nodes. iii) Our methods set the new state-of-the-art performance on smart contract vulnerability detection, and overall provide insights into the challenges and opportunities
• In contrast to existing methods, we explicitly model the fallback mechanism of smart contracts, consider rich dependencies between program elements, and explore the possibility of using novel graph neural networks for vulnerability detection
• Extensive experiments show that our method significantly outperforms state-of-theart methods and other neural networks

添加“小脉”微信，留言“IJCAI 2020”,即可加入【IJCAI 2020会议交流群】，与IJCAI 2020论文作者面对面沟通！