（知识图谱补全）Meta Relational Learning元关系学习

A Survey on Knowledge Graphs: Representation, Acquisition, and Applications
 

The long-tail phenomena exist in the relations of knowledge graphs. Meanwhile, the real-world
scenario of knowledge is dynamic, where unseen triples are usually acquired. The new scenario, called as meta relational learning or few-shot relational learning, requires models to predict new relational facts with only a very few samples
知识图谱的关系中存在长尾现象。而现实世界知识的情景是动态的，也时常获得不可见三元组。新的情景被称为元关系学或少样本学习，只是需要少量样本情形下，用模型预测新的关系事实。

Targeting at the previous two observations, GMatching [104] develops a metric based few-shot learning method with entity embeddings and local graph structures. It encodes one-hop neighbors to capture the structural information with R-GCN and then takes the structural entity embedding for multistep matching guided by long short-term memory (LSTM) networks to calculate the similarity scores. Meta-KGR [105], an optimization-based meta-learning approach, adopts model agnostic meta-learning for fast adaption and reinforcement learning for entity searching and path reasoning. Inspired by modelbased and optimization-based meta-learning, MetaR [106] transfers relation-specific meta information from support set to query set, and archives fast adaption via loss gradient of highorder relational representation. Zhang et al. [107] proposed joint modules of heterogeneous graph encoder, recurrent autoencoder, and matching network to complete new relational facts with few-shot references. Qin et al. [108] utilized GAN to generate reasonable embeddings for unseen relations under the zeroshot learning setting. Baek et al. [109] proposed a transductive meta-learning framework, called Graph Extrapolation Networks(GEN), for few-shot out-of-graph link prediction in knowledge graphs

针对前面两个观察，

GMatching开发了一种基于少量样本学习的度量方法，使用了是提前入和局部图结构。 通过R-GCN一跳的邻居获取结构信息，然后将结构实体嵌入到长短时记忆网络，计算相似度得分。

Meta-KGR一个基于优化的元学习方法，采用模型不可知元学习进行快速适应以及强化学习进行实体搜索和路径推理。

受基于模型和基于优化的元学习的启发，MetaR将关系特定的元信息从支持及转换到查询集，并通过高阶关系表示的损失梯度实现快速自适应

Zhang等人提出了异质图的联合模型，包括编码器、循环自编码器和匹配网络，通过小数据量引用补全新的关系事实。

Qin等人[108]利用GAN在零距离学习设置下为不可见的关系生成合理的嵌入

Baek等人[109]提出了一种转导元学习框架，称为图外推网络(GEN)，用于知识图谱中小数据量外图链接预测。