Along the Time: Timeline-traced Embedding for Temporal Knowledge Graph Completion

Along the Time: Timeline-traced Embedding for Temporal Knowledge Graph Completion

💕当前模型存在的问题💕

1.most existing models ignore the independence of semantic and temporal information（simply concatenate or multiply the semantic and temporal information together）
2.current models have difficulty distinguishing representations of the same entity or relaton at different timestamps

💕论文提出的方法💕

TimeLine-Traced Knowledge Graph Embedding method(TLT-KGE)
具体实现目标：embed the entities and relations with timestamps as a complex vector or a quaternion vector

💕相关工作💕

🐸Static Knowledge Graph Embedding🐸

1.Translation-based

具体做法：assume the relation embedding as a translation or rotation from the subject to the object
具体方法：
①TransE
②TransH
③TransR
④RotatE

2.Semantic matching

具体做法：compute the plausibility of a given fact via matching the semantic information of entities and relations
具体方法：
①DistMult：use a bilinear function to calculate the score
②ComplEx：extend DistMult to complex space
③QuatE：extend DistMult to quaternion elliptic space
④DualE：extend QuatE to dual vector space
⑤ConvE
⑥ConvKB
⑦Tucker：tucker-decomposition-based method

🐸Temporal Knowledge Graph Embedding🐸

具体方法：
①TA-TransE：incorporate the embeddings of time information to TransE
②TA-DistMult：incorporate the embeddings of time information to DistMult
③HyTE：extend TransH by projecting the entities and relations to a time-specific hyperplane
④DE-SimplE：introduce a diachronic part for entities and relations
⑤TeRo：inspired by RotatE,utilize a rotation operation on both subject and object to evaluate the semantic scores of a given fact
⑥T(NT)ComplEx：incorporate the time embedding into a three-order tensor and uses a novel regularization method to improve link prediction performance
⑦RotateQVS：model the temporal changes with rotation operation in quaternion vector space and uses a score function that is similar to TransE
⑧BoxTE：box embedding model for temporal knowledge graph completion, which is based on the static KGE model BoxE
⑨RE-NET
⑩RE-GCN

💕汉密尔顿四元数预备知识💕

1.a quaternion 𝑞 is consist of one real component and three imaginary components 𝑖, 𝑗, 𝑘, which is defines as 𝑞 = 𝑎 + 𝑏i + 𝑐j + 𝑑k,where 𝑎, 𝑏, 𝑐, 𝑑 are real numbers and i, j, k are imaginary units
2.it can also be viewed as a quadruple (𝑎, 𝑏, 𝑐, 𝑑)
3.Conjugate

4.Norm

5.Inner Product

6.Hamilton Product

💕论文具体方法💕

🐸概述🐸

一.divide the embedding of entities or relations into two independent parts
1.semantic meaning of entities or relations(semantic part)
2.timestamps(temporal part)
二.TLT-KGE is equipped with two additional components to strengthen the expression of timestamps and relations
1.🙌Shared time window🙌 : utilize shared embeddings for time windows to strengthen the connection of events that happen within a time scope
2.🙌Relation-timestamp composition🙌 : enhance the expression of a relation at specific time

🐸Proposed Method: TLT-KGE🐸

🤷‍♂️Timeline-tracedEmbedding🤷‍♂️

🤡Semantic Part🤡 :

🤡Temporal Part 🤡:

🤷‍♂️Combine to a Uniform Embedding🤷‍♂️

🤡Complex Vector Space🤡 :

🤡Quaternion Vector Space🤡 :

🤷‍♂️Score Functions🤷‍♂️

🤡Score Function of TLT-KGE (Complex)🤡 ：

🤡Score Function of TLT-KGE (Quaternion) 🤡：

🤷‍♂️Additional Modules🤷‍♂️

🤡Shared Time Window (STW) 🤡：

🤡Relation-Timestamp Composition (RTC) 🤡：

🐸Regularization🐸

🤡the semantic embeddings of entities and relations🤡 ：

🤡the temporal embeddings of timestamps🤡 ：

🐸Loss function🐸

💕实验💕

🐸实验数据集🐸

🐸baselines🐸

static KGE models:
①TransE ②DistMult
temporal KGE models：
①TTransE ②HyTE ③TA-DistMult ④DE-SimplE ⑤TeRo ⑥T(NT)ComplEx ⑦BoxTE

🐸评估标准🐸

1.Mean Rank ： MR, the mean of all the predicted ranks
2.Mean Reciprocal Rank ：MRR, the mean of all the reciprocals of predicted ranks
3.Hits@n ：the proportion of ranks not larger than n
Lower MR and larger MRR and Hits@𝑛 indicate better performance.

🐸实验结果🐸

🐸参数受限实验结果🐸

🐸消融实验🐸

🐸参数学习🐸