Search

Knowledge graph (KG) is a powerful tool in many areas, but it is impossible to take in all knowledge during construction for the complexity of relations among natural entities. In practical terms, many relations within KG have few samples to learn from, so few-shot KG completion matters. Existing approaches generally complete KGs by learning either static representations of entities and relations or the dynamic representations of fixed model parameters, while ignoring the importance of dynamic fine-tuning of model parameters for rapidly mining task-related information. In this paper, we propose a novel task-related network based on meta-learning (MTRN) for few-shot knowledge graph completion. Firstly, a task-related neighbor-aware encoder, which can distinguish the contributions of different neighbors, is designed to mine the task-related information of neighbors from the heterogeneous background graph and to enhance task entity representations. Secondly, a self-attention entity-pair encoder is proposed to generate the representations of references by modeling the interaction of head and tail entities. Then MTRN models references, that is entity pairs in the reference set, through a matching network to differentiate the contributions of different entity pairs to task relations and obtain the task-related representations with fine-grained semantics. When performing different knowledge graph completion tasks, MTRN can extract the features of few-shot references in a meta-learning way to dynamically fine-tune model parameters, making MTRN quickly adapt to diverse tasks in few-shot setting. The evaluation on two public datasets of knowledge graph completion shows that MTRN achieves competitive completion performance under different few-shot sizes. [ABSTRACT FROM AUTHOR]

Network embedding can effectively dig out potentially useful information and discover the relationships and rules which exist in the data, that has attracted increasing attention in many real-world applications. The goal of network embedding is to map high-dimensional and sparse networks into low-dimensional and dense vector representations. In this paper, we propose a network embedding method based on high-degree penalty and adaptive negative sampling (NEPS). First, we analyze the problem of imbalanced node training in random walk and propose an indicator base on high-degree penalty, which can control the random walk and avoid over-sampling high-degree neighbor node. Then, we propose a two-stage adaptive negative sampling strategy, which can dynamically obtain negative samples suitable for the current training according to the training stage to improve training effect. By comparing with seven well-known network embedding algorithms on eight real-world data sets, experiments show that the NEPS has good performance in node classification, network reconstruction and link prediction. The code is available at: https://github.com/Andrewsama/NEPS-master. [ABSTRACT FROM AUTHOR]

With the exponential growth of the volume of scientific literatures, academic evaluation is becoming one of the important problems of scientometrics. Considering the influence of disruption of papers, in this paper, we propose a disruption based PageRank (DPRank) model to rank the scientific significance of scientific papers. The disruption value of a paper is calculated by the local network structure of the paper in the citation network. DPRank model treats each citation differently based on paper's temporal information and disruption value. In DPRank model, the scores delivered from a paper's upstream nodes decay exponentially with the time gap between the publication time and reference time, and the decay rate is proportional to the paper's disruption value. We use the APS data set and some highly valued papers as benchmarks to evaluate the performance of the proposed DPRank model. The experimental results show that the DPRank model has a relatively better performance for both old and newly published papers, and can outperform other methods in identifying the benchmark papers. [ABSTRACT FROM AUTHOR]

In recent years, many attributednetwork have emerged, such as Facebook networks in social networks, protein networks and academic citation networks. In order to find communities where the nodes are tightly connected and have attributes similar to each other by unsupervised learning and improve the accuracy of community detection to make better analysis of the attributed networks, we propose a two-stage attributed network community detection combined with network embedding and parameter-free clustering. In the first stage, we build an attributed network embedding framework that integrates common neighbor information and node attributes. We define node similarity in terms of local link information, jointly model it with attribute proximity, and then adopt the distributed algorithm to obtain the embedding vector of each node. In the second stage, the number of communities can be decided automatically based on curvature and modularity, and the community detection results can be obtained by clustering the embeddings. The performance experiments of our method compared with some representative approaches are tested on real network datasets. The experimental results validate the effectiveness and superiority of our approach. [ABSTRACT FROM AUTHOR]