Search

Graph Attention Networks (GATs) and Graph Convolutional Neural Networks (GCNs) are two state-of-the-art architectures in Graph Neural Networks (GNNs). It is well known that both models suffer from performance degradation when more GNN layers are stacked, and many works have been devoted to address this problem. We notice that main research efforts in the line focus on the GCN models, and their techniques cannot well fit the GAT models due to the inherent difference between these two architectures. In GAT, the attention mechanism is limited as it ignores the overwhelming propagation from certain nodes as the number of layers increases. To sufficiently utilize the expressive power of GAT, we propose a new version of GAT named L ayer-wise S elf-adaptive GAT (LSGAT), which can effectively alleviate the oversmoothing issue in deep GAT and is strictly more expressive than GAT. We redesign the attention coefficients computation mechanism adaptively adjusted by layer depth, which considers both immediate neighboring and non-adjacent nodes from a global view. The experimental evaluation confirms that LSGAT consistently achieves better results on node classification tasks over relevant counterparts. • We explore and introduce the challenge of overwhelming propagation in deep GATs. • The novel deep GAT model LSGAT is introduced to mitigate the oversmoothing problem. • Extensive experiments demonstrate the superiority of our proposed model. [ABSTRACT FROM AUTHOR]

This work investigates a novel m -closest entity (m CE) matching problem over geographic heterogeneous information networks (Geo-HINs). That is, given a Geo-HIN G and m query graphs { q 1 , q 2 , ... , q m } , m CE matching aims to find a group of geographic entities (geo-entities) whose patterns match the query graphs { q 1 , q 2 , ... , q m } correspondingly, for which the maximum distance between any geo-entity pair (i.e., the diameter) in the group is minimized. As a fundamental problem, the m CE matching can be applied for many scenarios, e.g., travel itinerary recommendation and city planning. The existing works have not simultaneously considered the characteristics of patterns matching and spatial search so that they cannot solve our problem, which is computationally expensive. To solve this problem efficiently, we propose a unified framework named F u z z y − E x a c t framework to process entity matching and spatial search comprehensively, in which pruning abilities at non-spatial and spatial layers are cooperatively explored. Two mutually adaptive auxiliary data structures named A r c − T r e e and A r c − F o r e s t are devised to maintain the intermediate search results which are exploited to enhance the search process between non-spatial and spatial layers. Experimental results demonstrate that our algorithm can outperform the baseline methods by 2 orders of magnitude on runtime. • A novel m -closest entity matching problem over HINs is devised. • Pruning abilities at both the non-spatial and spatial layers are explored. • Our approach can achieve much better performance than the baseline methods. [ABSTRACT FROM AUTHOR]

Abstract: A process trace describes the steps taken in a workflow to generate a particular result. Understanding a process trace is critical to be able to verify data, enable its re-use and to make appropriate decisions. Given many process traces, each with a large amount of very low level information, it is a challenge to make process traces meaningful to different users. It is more challenging to compare two complex process traces generated by heterogeneous systems and having different levels of granularity. In this paper, we present a novel notion of multi-granularity process trace that attempts to capture the conceptual abstraction of large process traces at different levels of granularity by leveraging ontology information. Based on this notion, graph matching based algorithms with semantic filtering are developed to efficiently and effectively compute the similarity between two process traces by considering both structural similarity and semantic similarity. Our experiment using both real world and synthetic datasets demonstrates that our techniques provide a practical approach for process trace similarity measurement. [Copyright &y& Elsevier]

Knowledge Graph Question Answering (KGQA) is a major branch of question answering tasks, which can answer fact questions effectively by using the reasonable characteristics of the knowledge graph. Currently, lots of related works combined with a variety of deep learning models are presented for the KGQA task. However, there are still some challenges, such as topic entity recognition under ambiguity expression, semantic level representation of natural language, efficient construction of searching space for answers, etc. In this paper, we propose a comprehensive approach for complex question answering over KG. Firstly, during the stage of topic entity recognition, a deep transition model is constructed to extract topic entities, and an efficient entity linking strategy is presented, which combines character matching and entity disambiguation model. Secondly, for candidate path ranking, a dynamic candidate path generation algorithm is proposed to efficiently create the candidate answer set. And four dedicated similarity calculation models are designed to handle the intricate condition of complex questions with long sequence and diversity expression. Moreover, a fusion policy is proposed to make decision for the final correct answer. We evaluate our approach on CKBQA, a Chinese knowledge base question answering dataset, from CCKS2019 competition. Experimental results demonstrate that the improvements in each process are effective and our approach achieves better performance than the best team in CCKS2019 competition. [ABSTRACT FROM AUTHOR]