【论文阅读】Unifying Knowledge Graph Learning and Recommendation

Unifying Knowledge Graph Learning and Recommendation: Towards a Better Understanding of User Preferences

总体大纲如下图

记录下：9月6号

组合两个模型：TransH和TUP
TransH：考虑关系和实体的1对N，N对1，N对N的情况，实现kg的补全增强
TUP：包括偏好引入和超平面映射

偏好归纳组件
给定user-item对（u,i），组件从latent factors集合P中归纳出一种偏好。
两种策略：
Hard Strategy：假设当用户对项目做出决定时，只有一个偏好生效
soft Strategy：注意力机制融合多个偏好
超平面映射
和TransH思想类似，user-item对看作h，t，用户对item的偏好看作r，用偏好p代替kg中的r.

KTUP: 组合上面两个模型
同时提出加强item embedding，perference embedding

   i' = i + e;p' = p + rw_p' = w_p + w_r

然后进行超平面映射，得到新的打分函数

问题：知识图谱的补全增强体现在哪里？TransH算法进行知识图谱的补全

记录下：9月7号，学习TransH知识图谱补全
参考：
transH算法实现知识图谱补全实验

TransH认为：TransE算法在处理自反关系，多对一，一对多，多对多关系时，会使不同的实体拥有相同的嵌入（h+r=t,h+r=t’=>t=t’)
提出了超平面，将头实体和尾实体映射到同一个超平面上，dr是两者在超平面上的平移

知识图谱补全：链接预测
通过TransH建模后，得到了每个实体和关系的嵌入向量，利用嵌入向量，可以进行知识图谱的链接预测
将三元组(head,relation,tail)记为(h,r,t)

链接预测分为三类：
头实体预测：(?,r,t)
关系预测：(h,?,t)
尾实体预测：(h,r,?)
但原理很简单，利用向量的可加性即可实现。以(h,r,?)的预测为例：

假设t’=h+r，则在所有的实体中选择与t’距离最近的向量，即为t的的预测值

问题： $w_p'$ 怎么获得

jTransUP.py perference归纳代码

# u_e, i_e : batch * dim or batch * item * dim
def getPreferences(self, u_e, i_e, use_st_gumbel=False):# use item and user embedding to compute preference distribution# pre_probs: batch * rel, or batch * item * rel# p = p+r# torch.t转置# pre_probs是相似度分数，不清楚为啥/2pre_probs = torch.matmul(u_e + i_e, torch.t(self.pref_embeddings.weight + self.rel_embeddings.weight)) / 2if use_st_gumbel:# ST Gumbel采样，获得一个热矢量pre_probs = self.st_gumbel_softmax(pre_probs)# 注意力参数和相似度得分成正比r_e = torch.matmul(pre_probs, self.pref_embeddings.weight + self.rel_embeddings.weight) / 2# norm就是w_p',pref_norm_embeddings像是正则化后的pref_embeddingnorm = torch.matmul(pre_probs, self.pref_norm_embeddings.weight + self.norm_embeddings.weight) / 2return pre_probs, r_e, norm

pref_norm_embeddings的初始化

pref_norm_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
nn.init.xavier_uniform(pref_norm_weight)
self.pref_norm_embeddings = nn.Embedding(self.rel_total, self.embedding_size)
self.pref_norm_embeddings.weight = nn.Parameter(pref_norm_weight)
normalize_pref_norm_emb = F.normalize(self.pref_norm_embeddings.weight.data, p=2, dim=1)
self.pref_norm_embeddings.weight.data = normalize_pref_norm_emb

根据norm超平面映射

_, r_e, norm = self.getPreferences(u_e, ie_e, use_st_gumbel=self.use_st_gumbel)proj_u_e = projection_transH_pytorch(u_e, norm)
proj_i_e = projection_transH_pytorch(ie_e, norm)def projection_transH_pytorch(original, norm):return original - torch.sum(original * norm, dim=len(original.size())-1, keepdim=True) * norm

【论文阅读】Unifying Knowledge Graph Learning and Recommendation相关推荐

Unifying Task-oriented Knowledge Graph Learning and Recommendation
Unifying Task-oriented Knowledge Graph Learning and Recommendation ABSTRACT 背景将知识图纳入推荐系统(知识感知推荐),以辅 ...
论文阅读 Enhancing knowledge graph embedding with relational constraints
Enhancing knowledge graph embedding with relational constraints 利用关系约束增强知识图嵌入发表于:Neurocomputing 429 ...
论文阅读《Knowledge Graph Refinement: A Survey of Approaches and Evaluation Methods》
论文链接一.谷歌的知识图谱谷歌的知识图谱是在2012年向公众推出的,也是在这个时候,"知识图谱"这个术语被创造出来. 谷歌本身对于知识图谱的构建是相当保密的:只有少数外部来源讨 ...
【论文阅读】Spatio-Temporal Graph Convolutional Networks:...Traffic Forecasting[时空图卷积网络:用于交通预测的深度学习框架]（1）
[论文阅读]Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecastin ...
论文阅读|node2vec: Scalable Feature Learning for Networks
论文阅读|node2vec: Scalable Feature Learning for Networks 文章目录论文阅读|node2vec: Scalable Feature Learning ...
论文阅读笔记：Geography-Aware Sequential Location Recommendation
论文阅读笔记:Geography-Aware Sequential Location Recommendation 文章目录论文阅读笔记:Geography-Aware Sequential Loc ...
论文阅读笔记：SCAN: Learning to Classify Images without Labels
论文阅读笔记:SCAN: Learning to Classify Images without Labels 摘要简介和相关工作方法表征学习语义聚类损失 2.3 通过自标记进行微调 3 实验 ...
Zero-shot Learning零样本学习论文阅读（一）——Learning to detect unseen object classes by between-class attribute
Zero-shot Learning零样本学习论文阅读(一)--Learning to detect unseen object classes by between-class attribute ...
年龄论文阅读——Deep Label Distribution Learning With Label Ambiguity
论文阅读--Deep Label Distribution Learning With Label Ambiguity 版权声明:本文为博主原创文章,未经博主允许不得转载.https://blog.c ...

【论文阅读】Unifying Knowledge Graph Learning and Recommendation

Unifying Knowledge Graph Learning and Recommendation: Towards a Better Understanding of User Preferences

【论文阅读】Unifying Knowledge Graph Learning and Recommendation相关推荐

最新文章

热门文章