Neural Graph Collaborative Filtering

Wang X, He X, Wang M, et al. Neural Graph Collaborative Filtering[J]. arXiv preprint arXiv:1905.08108, 2019.

论文的主要贡献是提出了一种embeding propagation的方式，能够利用high order范围内的实体，训练得到用户和物品的embeding。结合知识图谱做推荐。

1 背景

ONN(Operation-aware Neural Networks for user response)是2018年腾讯广告算法比赛最优的推荐算法。主要任务是预测用户点击推荐广告的概率(click-through rate, CTR)或者进行其它期望的行为(conversion rate, CVR)。在基本的通用的Base model上，将PNN与FFM结合起来，实现了在embedding层的每一个feature对于不同operation（内积或者外积）有不同的表示，之后进入MLP，得到更好的预测结果

The intuitions behind Tree-structured Parzen estimator

TPE：一种基于贝叶斯推断的超参数调优方法。

HOW POWERFUL ARE GRAPH NEURAL NETWORKS?

This paper:

1. characterize how expressive different GNN variants are in learning to represent and distinguish between different graph structures
2. show that GNNs are at most as powerful as the WL test in distinguishing graph structures.
3. identify graph structures that cannot be distinguished by popular GNN variants, such as GCN (Kipf & Welling, 2017) and GraphSAGE (Hamilton et al., 2017a)
4. develop a simple neural architecture, Graph Isomorphism Network (GIN)

本文就提出了一种无监督的方法。核心思想：通过特定的游走方式进行采样，对于每个点都会生成 对应的序列。再将这些序列视为文本导入skip-gram模型，即可得 到每个节点的向量

Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs

本文讨论了graph的heterophily

A Three-Way Model for Collective Learning on Multi-Relational Data

2011

we propose the relational learning approach RESCAL which is based on a tensor factorization that is related to DEDICOM but does not exhibit the same constraints.

Convolutional 2D Knowledge Graph Embeddings

2018-7-4

1 Introduction

第一个利用CNN学习KGE的方法。

A Novel Embedding Model for Knowledge Base Completion Based on Convolutional Neural Network

2018-3-13 ConvKB

Adaptive Convolution for Multi-Relational Learning

2019-6

1 Introduction

Learning with multi-relational data plays a pivotal role in many application domains, ranging from social networks or recommender systems to large-scale knowledge bases (KBs)

ConvR的思想是从relation中构造filter，然后卷积于subject embedding，最后投影，与object embedding做点积。

这样的做法就导致了ConvR的另一个优势，减少了参数的数量。